Efficiently Sampling Interval Patterns from Numerical Databases
Extracted source text · arXiv 2512.00105
%%
%% Copyright 2007-2025 Elsevier Ltd
%%
%% This file is part of the 'Elsarticle Bundle'.
%% ---------------------------------------------
%%
%% It may be distributed under the conditions of the LaTeX Project Public
%% License, either version 1.3 of this license or (at your option) any
%% later version. The latest version of this license is in
%% http://www.latex-project.org/lppl.txt
%% and version 1.3 or later is part of all distributions of LaTeX
%% version 1999/12/01 or later.
%%
%% The list of all files belonging to the 'Elsarticle Bundle' is
%% given in the file `manifest.txt'.
%%
%% Template article for Elsevier's document class `elsarticle'
%% with numbered style bibliographic references
%% SP 2008/03/01
%% $Id: elsarticle-template-num.tex 272 2025-01-09 17:36:26Z rishi $
%%
\documentclass[preprint,12pt]{elsarticle}
%% Use the option review to obtain double line spacing
%% \documentclass[authoryear,preprint,review,12pt]{elsarticle}
%% Use the options 1p,twocolumn; 3p; 3p,twocolumn; 5p; or 5p,twocolumn
%% for a journal layout:
%% \documentclass[final,1p,times]{elsarticle}
%% \documentclass[final,1p,times,twocolumn]{elsarticle}
%% \documentclass[final,3p,times]{elsarticle}
%% \documentclass[final,3p,times,twocolumn]{elsarticle}
%% \documentclass[final,5p,times]{elsarticle}
%% \documentclass[final,5p,times,twocolumn]{elsarticle}
%% For including figures, graphicx.sty has been loaded in
%% elsarticle.cls. If you prefer to use the old commands
%% please give \usepackage{epsfig}
%% The amssymb package provides various useful mathematical symbols
\usepackage{amssymb}
\usepackage{float}
%% The amsmath package provides various useful equation environments.
\usepackage{amsmath}
%% The amsthm package provides extended theorem environments
\usepackage{amsthm}
\usepackage{todonotes}
\usepackage{xurl}
\usepackage[ruled,linesnumbered, vlined]{algorithm2e}
%% The lineno packages adds line numbers. Start line numbering with
%% \begin{linenumbers}, end it with \end{linenumbers}. Or switch it on
%% for the whole article with \linenumbers.
%% \usepackage{lineno}
\usepackage{xcolor}
\journal{Data \& Knowledge Engineering}
\input{cmd}
% \input{tex/CoverLetter}
% \coverletter
\begin{document}
\begin{frontmatter}
%% Title, authors and addresses
%% use the tnoteref command within \title for footnotes;
%% use the tnotetext command for theassociated footnote;
%% use the fnref command within \author or \affiliation for footnotes;
%% use the fntext command for theassociated footnote;
%% use the corref command within \author for corresponding author footnotes;
%% use the cortext command for theassociated footnote;
%% use the ead command for the email address,
%% and the form \ead[url] for the home page:
%% \title{Title\tnoteref{label1}}
%% \tnotetext[label1]{}
%% \author{Name\corref{cor1}\fnref{label2}}
%% \ead{email address}
%% \ead[url]{home page}
%% \fntext[label2]{}
%% \cortext[cor1]{}
%% \affiliation{organization={},
%% addressline={},
%% city={},
%% postcode={},
%% state={},
%% country={}}
%% \fntext[label3]{}
\title{
Efficiently Sampling Interval Patterns\\ from Numerical Databases}
%DB: Addressing the Long Tail Phenomenon in Interval Pattern Sampling"
%fficient Frequency based Interval Patterns Sampling
%% use optional labels to link authors explicitly to addresses:
%% \author[label1,label2]{}
%% \affiliation[label1]{organization={},
%% addressline={},
%% city={},
%% postcode={},
%% state={},
%% country={}}
%%
%% \affiliation[label2]{organization={},
%% addressline={},
%% city={},
%% postcode={},
%% state={},
%% country={}}
\author[unicaen]{Djawad Bekkoucha} %% Author name
\author[EPITA]{Lamine Diop}
\author[unicaen]{Abdelkader Ouali}
\author[unicaen]{Bruno Crémilleux}
\author[unicaen]{Patrice Boizumault}
%% Author affiliation
\affiliation[unicaen]{
organization={Université de Caen Normandie, ENSICAEN, CNRS, Normandie Univ, GREYC UMR6072},
addressline={6 Boulevard Marechal Juin},
city={Caen},
postcode={F-14000},
country={France}
}
\affiliation[EPITA]{
organization={EPITA Research Laboratory (LRE)},
addressline={Le Kremlin-Bicetre},
city={Paris},
postcode={FR-94276},
country={France}
}
%% Abstract
\begin{abstract}
%% Text of abstract
Pattern sampling has emerged as a promising approach for information discovery in large databases, allowing analysts to focus on a manageable subset of patterns. In this approach, patterns are randomly drawn based on an interestingness measure, such as frequency or hyper-volume. This paper presents the first sampling approach designed to handle interval patterns in numerical databases. This approach, named \fips{}, samples interval patterns proportionally to their frequency. It uses a multi-step sampling procedure and addresses a key challenge in numerical data: accurately determining the number of interval patterns that cover each object. We extend this work with \hips{}, which samples interval patterns proportionally to both their frequency and hyper-volume. These methods efficiently tackle the well-known long-tail phenomenon in pattern sampling. We formally prove that \fips{} and \hips{} sample interval patterns in proportion to their frequency and the product of hyper-volume and frequency, respectively. Through experiments on several databases, we demonstrate the quality of the obtained patterns and their robustness against the long-tail phenomenon.
\end{abstract}
%%Graphical abstract
%\begin{graphicalabstract}
%\includegraphics{grabs}
%\end{graphicalabstract}
%%Research highlights
%\begin{highlights}
%\item Research highlight 1
%\item Research highlight 2
%\end{highlights}
%% Keywords
\begin{keyword}
%% keywords here, in the form: keyword \sep keyword
Numerical data \sep Data mining \sep Pattern sampling \sep Interval patterns.
%% PACS codes here, in the form: \PACS code \sep code
%% MSC codes here, in the form: \MSC code \sep code
%% or \MSC[2008] code \sep code (2000 is the default)
\end{keyword}
\end{frontmatter}
\input{tex/Introduction}
\input{tex/Preliminaries}
\input{tex/Relatedworks}
\input{tex/contribution}
\input{tex/Experiments}
\input{tex/ConclusionPerspectives}
%% The Appendices part is started with the command \appendix;
%% appendix sections are then done as normal sections
\appendix
\input{tex/Appendix_proofs}
\input{tex/Appendix_experimentalResults}
%% If you have bib database file and want bibtex to generate the
%% bibitems, please use
%%
%% \bibliographystyle{elsarticle-num}
%% \bibliography{<your bibdatabase>}
%% else use the following coding to input the bibitems directly in the
%% TeX file.
%% Refer following link for more details about bibliography and citations.
%% https://en.wikibooks.org/wiki/LaTeX/Bibliography_Management
\bibliographystyle{elsarticle-num}
\bibliography{biblio}
%\begin{thebibliography}{00}
%% For numbered reference style
%% \bibitem{label}
%% Text of bibliographic item
%\end{thebibliography}
\end{document}
\endinput
%%
%% End of file `elsarticle-template-num.tex'.
\newcommand{\ouali}[1]{{\color{blue}#1}}
\newcommand{\hips}[0]{\textsc{HFips}}
\newcommand{\fips}[0]{\textsc{Fips}}
\newcommand{\Cfips}[0]{\textsc{CFips}}
\newcommand{\alea}[0]{\textsc{Uniform}}
\newcommand{\NIP}[0]{NIP}
\newcommand{\IH}[1]{IPH(#1)}
\newcommand{\TIL}[2]{TIL_{#1}(#2)}
\newcommand{\I}[1]{I(#1)}
\newcommand{\J}[1]{J(#1)}
\newcommand{\TermU}{U}
\newcommand{\TermA}{A}
\newcommand{\TermB}{B}
\newcommand{\TermT}{T}
\newcommand{\TermS}{S}
\newcommand{\TermH}{H}
%NIP terms
\newcommand{\PUC}[4]{P^{U_{{#1}^{#2}_{#3}}}_{\val{}{}}(#4)}
\newcommand{\PAC}[4]{P^{A_{{#1}^{#2}_{#3}}}_{\val{}{}}(#4)}
\newcommand{\PSC}[4]{P^{S_{{#1}^{#2}_{#3}}}_{\val{}{}}(#4)}
\newcommand{\PHC}[4]{P^{H_{{#1}^{#2}_{#3}}}_{\val{}{}}(#4)}
\newcommand{\PTC}[4]{P^{T_{{#1}^{#2}_{#3}}}_{\val{}{}}(#4)}
%Example of constraints:
\newcommand{\INC}[3]{Inc^{#1}_{#2}(#3)}
\newcommand{\EXC}[3]{Exc^{#1}_{#2}(#3)}
\newcommand{\SUP}[3]{Sup^{#1}_{#2}(#3)}
\newcommand{\SUPEQ}[3]{SupEq^{#1}_{#2}(#3)}
\newcommand{\INF}[3]{Inf^{#1}_{#2}(#3)}
\newcommand{\INFEQ}[3]{InfEq^{#1}_{#2}(#3)}
\newtheorem{definition}{Definition}
\newtheorem{theorem}{Theorem}
\newtheorem{property}{Property}
%new commands
%\newcommand{\todo}[2]{{\color{red} {\bf todo}[#1][\\#2]}}
\newcommand{\locked}[2]{{\color{blue} {\bf locked}[#1][\\#2]}}
\newcommand{\tred}[1]{{\color{red} #1}}
\newcommand{\tblue}[1]{{\color{blue} #1}}
% math stuff
% \newtheorem{theorem}{Theorem}[]
% \newtheorem{corollary}{Corollary}[theorem]
% \newtheorem{lemma}[theorem]{Lemma}
\newtheorem{example}{Example}
\newtheorem{prop}{Proposition}
% \newtheorem{definition}{Definitoin}[]
\def\uniquevalues{$\frac{1}{|\mathcal{M}|} \sum_{m=1}^{|\mathcal{M}|} \frac{|{\text{distinct values}}_m|}{|\mathcal{G}|} $}
\newtheorem{Definition}{Définition}
\newtheorem{Proposition}{Proposition}
\let\oldnl\nl% Store \nl in \oldnl
\newcommand{\nonl}{\renewcommand{\nl}{\let\nl\oldnl}}% Remove line number for one line
%symblos
%ips
\def\N{\mathcal{N}} % the numerical dataset
%\def\B{\mathcal{B}} % the binary dataset
\def\M{\mathcal{M}} % the set of attributes
\def\im{{m}} % the set of attributes
\def\L{\mathcal{L}} % the set of binary attributes
\def\G{\mathcal{G}} % the set of objects
\def\ig{{g}} % the index of objects
\def\io{{o}} % the second index of objects
\def\R{\mathcal{R}} % the relation
\def\W{\mathcal{N}} % the set of values in \N
\def\IP{\mathcal{V}} % the set of values in \N
\def\BP{\mathcal{P}} % the set of values in \N
\def\BE{G} % the extension of binary pattern
\def\BI{M} % the intention of binary pattern
\def\PSD{D} % the set of description of patterns
\def\gAsIP{\mathcal{B}}
%measure
\def\cover{{\it cover}} % the set of values in \N
\def\freq{{\it freq}} % the set of values in \N
\def\vol{{\it vol}} % the set of values in \N
%cp
\def\vars{\mathcal{V}}
\def\doms{\mathcal{D}}
\def\cons{\mathcal{C}}
\newcommand{\inff}[1]{\underline{#1}}
\newcommand{\supp}[1]{\overline{#1}}
%\def\colival{{H}}
\def\icptext{\textsc{IntervalClosedPattern}}
\def\cpcip{\textsc{cp4cip}}
\def\gccip{\textsc{gc4cip}}
\def\gccluster{\textsc{GC4CLUSTER}}
\def\closedpattern{\textsc{closedpattern}}
\def\kmeans{\textsc{K-Means}}
\def\kgccip{\textsc{K-gc4cip}}
\def\cpim{\textsc{cp4im}}
\def\cphp{\textsc{cp4hp}}
\def\minintchange{\textsc{MinIntChange}}
\def\lcm{\textsc{LCM}}
%sampling
\def\lang{\mathcal{L}} % language of patterns
\def\langC{\mathcal{L}_{\cons}} % language of patterns under constraints
\def\cip{CIP}
\def\fcip{FCIP}
\def\ipscon{\textsc{IntPatStr}}
\newcommand{\lbX}[1]{\underline{x}_{#1}}
\newcommand{\ubX}[1]{\overline{x}_{#1}}
\newcommand{\minDomLBX}[1]{min({\underline{x}_{#1}})}
\newcommand{\maxDomLBX}[1]{max({\underline{x}_{#1}})}
\newcommand{\minDomUBX}[1]{min({\overline{x}_{#1}})}
\newcommand{\maxDomUBX}[1]{max({\overline{x}_{#1}})}
\newcommand{\ID}[1]{$ID^{\uparrow}_{#1}$}
\newcommand{\List}[2]{L^{#1}_{#2}}
\newcommand{\val}[2]{v_{\ig^{#1},\im^{#2}}}
\newcommand{\Connus}[2]{join ({#1},{#2})}
\newcommand{\Possibles}[2]{\delta({#1},{#2})}
%\newcommand{\Proj}[2]{S_{#1,#2}^{\leq}}
%\newcommand{\Projgeq}[2]{S_{#1,#2}^{\geq}}
\newcommand\concat{\mathbin{+\mkern-20mu+}}
%\newcommand{\concat}{++}
\def\close{{\it close}}
\def\cover{{\it cover}} % the set of values in \N
\def\freq{{\it freq}} % the set of values in \N
\def\desc{{\it desc}}
\section{Introduction}
Data scientists have a central role for knowledge discovery from data.
In practice, analysts want to interact (visualise, select, explore) not only with the data, but also with the patterns or models supported by the data. To carry out such processes, it is essential to produce high quality results within a very short time so that the analyst does not have to wait. Otherwise, there is a risk that they will disconnect from the system. Pattern sampling is one solution to this challenge~\cite{DBLP:conf/pakdd/DzyubaL17}.
Pattern sampling aims at randomly selecting a pattern with a probability proportional to an interestingness measure~\cite{DBLP:journals/pvldb/alhassan2009, DBLP:journals/datamine/DzyubaLR17, DBLP:conf/kdd/BoleyLPG11} , such as frequency. For example, a pattern $\IP_1$ that is twice as frequent as a pattern $\IP_2$ will be twice as likely to be selected.
A naive approach to pattern sampling is to generate the entire set of patterns and then perform a weighted draw based on the interestingness measure. However, in practice, this fails due to the large size of the search space, even for simple pattern languages such as itemsets, i.e. data described by Boolean values. An answer to this problem was provided by Boley et al.~\cite{DBLP:conf/kdd/BoleyLPG11}, who defined an approach based on two successive draws. Once the problem has been properly decomposed, this approach guarantees an exact draw proportionally to the distribution resulting from the interestingness measure. Itemsets are the patterns considered in~\cite{DBLP:conf/kdd/BoleyLPG11}.
Numerical data are present in a wide range of applications such as transcriptome analysis~\cite{DBLP:journals/isb/BlachonPBRBG07}, medicine~\cite{DBLP:journals/isb/BlachonPBRBG07} and energy~\cite{DBLP:conf/dsaa/TschoraGPPR23}.
%Numerical data are present in a wide range of applications such as transcriptome analysis, demographic data and agronomy for studying heat stress in cattle.
Even if handling numerical data to define patterns goes back to the origins of data mining~\cite{DBLP:conf/sigmod/SrikantA96}, pattern mining in numerical data remains a challenging task. A simple approach to cope with numerical data is to reuse existing methods for discrete data by first converting data into a binary representation~\cite{Dougherty95supervisedand}. However, it is well-known that the binarization process often leads to a loss of information.
Kaytoue at al.~\cite{kaytoue-ijcai11} introduced interval patterns, which preserve all the original information. However, this framework faces scalability issues due to the large number of generated patterns, which highlights the importance of pattern sampling in this context. To the best of our knowledge, there is only one sampling method designed for numerical data~\cite{DBLP:conf/sdm/GiacomettiS18}. This method uses a metric to construct neighbourhood patterns whose relevance is characterized by a density measure. Data are considered continuous and the method requires the setting of a parameter that defines the size (diameter) of a pattern.
{Our work} focuses on sampling in numerical data. We {strive} to keep the complete original information expressed by the numerical data by using \textit{Interval Patterns}~\cite{kaytoue-ijcai11}. We present \fips{} and \hips{}, two exact and non-enumerative methods for {sampling interval patterns}. \fips{} samples interval patterns proportionally to their frequency, while \hips{} extends this by incorporating hyper-volume, resulting in patterns sampled proportionally to the product of their frequency and hyper-volume. The key challenge for both methods is to compute, for each object in the database, its contribution to the overall sampling distribution. For example, in the case of the frequency, the main challenge is to determine the exact number of interval patterns covering an object and without enumerating all covering patterns.
This allows us to bias the sampling toward patterns with high interestingness scores. We formally prove that \fips{} and \hips{} sample patterns proportionally to the desired distributions. Experimentally, we evaluate the quality of the obtained patterns by using several criteria (frequency, hyper-volume times frequency, impact of the long tail phenomenon, diversity, speed, and plausibility). \fips{} and \hips{} are the first methods for interval pattern sampling.
The paper is organized as follows. Section~\ref{sec:context} introduces the notations, definitions, and the problem statement. Section~\ref{sec:rw} discusses related work on pattern sampling. Section~\ref{sec:contrib} presents the \fips{} method, while Section~\ref{contrib:HFIPS} extends it with \hips{}. Finally, Section~\ref{sec:exp} reports an experimental evaluation of the proposed approaches.
\section{Preliminaries}
\label{sec:context}
% We introduce in this section notation and basic concepts we use in the paper.
%\subsection{Interval Pattern Mining}
\subsection{{Numerical Dataset}}
A {\it numerical dataset} $\N$ is defined by
a set of objects $\G$ where each object is described by a set of attributes $\M$.
Each attribute $m \in \M$ has a range $\W_m$ which {is a finite set containing all the values of the data occurring in attribute $m$}. %in $\N$.
An object $\ig \in \G$ is defined by a vector of numerical values $< \val{}{} >_{\forall \im \in\M}$.
%
A dataset where the values of all attributes are binary $\W_m=\{0,1\}, \forall m \in \M$, is a special case of a numerical dataset and referred as a {\it binary dataset}.
\begin{example}
Table~\ref{dataset:nv:exemple} shows a running example of a numerical dataset containing 5~objects $\G=\{g_1,g_2, g_3, g_4, g_5\}$,
each object is described by 3~attributes $\M=\{m_1,m_2, m_3\}$.
\end{example}
\begin{table}[]
\centering
\scalebox{1}{
\begin{tabular}{c c c c}
& $m_1$ & $m_2$ & $m_3$ \\
\hline
$g_1$&2 &8 &130\\
$g_2$&4 &12 &102\\
$g_3$&3 &7 &91\\
$g_4$&2 &9 &101\\
$g_5$&6 &12 &110\\
\end{tabular}
}
\caption{A running example of a numerical dataset $\N$}
\label{dataset:nv:exemple}
\end{table}
\subsection{{Interval Patterns}}
Patterns in numerical datasets can be represented in many ways, we use the notion of Interval Pattern~\cite{kaytoue-ijcai11}
\label{def:intervalpatterns}
which is defined as a vector of intervals $\IP = \langle [a_m, b_m] \rangle_{\forall m \in \M}$, where $a_m, b_m \in \W_m$ and $a_m \leq b_m$.
Each dimension of the vector $\IP$ corresponds to an attribute following a canonical order on the set of attributes $\M$. We denote $\gAsIP[g] = \langle [\val{}{}, \val{}{}] \rangle_{ \forall \im \in \M}$ as the vector of intervals corresponding to an object identified by $\ig$. An object $\ig$ is an occurrence of the interval pattern $\IP$ if each interval in the vector $\gAsIP[g]$ is included in the interval of $\IP$, i.e. $\gAsIP[g] \sqsubseteq \IP \iff [\val{}{}, \val{}{}] \subseteq [a_m, b_m], \forall \im \in \M$.
The cover of $\IP$ in $\N$ is the set of objects $\ig \in \G$ occurring in $\IP$, i.e. {$\cover(\IP)= \{\ig \in \mathcal{G} ~|~ \gAsIP[g] \sqsubseteq \IP\}$.}
\label{def:cover}
\begin{example}
{In the example dataset of Table \ref{dataset:nv:exemple}}, $\IP = \langle [3, 4], [7, 12], [91, 130] \rangle $ is an interval pattern covering the objects $\{g_2, g_3\}$. $\gAsIP[g_2]$$ = $ $\langle [4, 4]$, $[12, 12],$ $[102,102] \rangle$ is the vector of intervals identified by the object $g_2$ and an occurrence of $\IP$.
\end{example}
\label{def:frequency}
The frequency of $\IP$ is the {cardinality} of its cover, i.e. $\freq(\IP)=|\cover(\IP)|$.
Given a minimum frequency threshold $\theta$, the interval pattern $\IP$ is frequent if and only if $\freq(\IP) \geq \theta$.
\label{def:description}
{The smallest description of a subset of objects $G \subseteq \G$ is the smallest interval pattern covering the set of objects $G$}. Formally the smallest description of $G$ is the interval pattern $\IP$ {such that for each} $\ig ~\in~ G$, $\ig$ is an occurrence of $\IP$, i.e. $\desc(G)= \langle[a_m,~b_m]\rangle_{ \forall \im \in \M} \text{ such that } a_m = min(\{\val{}{}~|~ g~ \in G \}) \text{ and } b_m = max(\{\val{}{}~|~ g~ \in G \})$.
The hyper-volume of an interval pattern $\IP = \langle[a_m,~b_m]\rangle_{ \forall \im \in \M} $ corresponds to the product of the lengths of its intervals, i.e.
$\vol(\IP)= \prod_{[a_m,b_m] \in \IP} ( b_m - a_m )$. %{The density of an interval pattern $\IP$ is the ratio between its frequency and its hyper-volume, i.e.
%$density(\IP)= \frac{\freq(\IP)}{\vol(\IP)}$.}
{Note that for an interval pattern $\IP$, if the interval associated to an attribute $m \in \M$ has its lower bound equal to the smallest possible value of $\W_m$ and its upper bound equal to the highest possible value of $\W_m$, then the interval is unconstrained. Otherwise, it is considered constrained on the subset of values it contains.}
Let $\mathcal{L}$ be the language of interval patterns, which corresponds to the set of all possible interval patterns. The size of the search space is the product of the total number of possible intervals for each attribute. Formally this is given by:
\[
\prod_{m \in \M}{\sum_{k=1}^{|\W_m|}{k}} = \prod_{m \in \M}{\frac{|\W_{m}| (|\W_{m}|+1)}{2}}
\]
Consider the dataset presented in Table~\ref{dataset:nv:exemple}, which contains 5 objects and 3 attributes. Each of the attributes $m_1$ and $m_2$ contains 4 distinct values, respectively $\W_{m_1}=\{2,3,4,6\}$ and $\W_{m_2}=\{7,8,9,12\}$, and attribute $m_3$ contains 5 distinct values, namely $\W_{m_3}=\{91,101,102,110,130\}$. The total number of interval patterns in the search space is given by :
\[
\underbrace{\frac{4 \times 5}{2} }_{m_1} \times \underbrace{\frac{4 \times 5}{2}}_{m_2} \times \underbrace{\frac{5 \times 6}{2}}_{m_3} = 1500 \text{ interval patterns}
\]
This example illustrates the size of the interval pattern search space. An exhaustive enumeration of all these patterns would be infeasible both in terms of time and memory. Moreover, given the large number of patterns, identifying the relevant ones would become a challenging task. This motivates the use of pattern sampling to explore the search space more efficiently.
\subsection{Problem Statement}
Let $\Omega$ be a population and $f : \Omega \longrightarrow [0,1]$ a measure. The notation $x \sim f(\Omega)$ {denotes} that the element $x$ is drawn randomly from $\Omega$ with a probability distribution $\pi(x) = f(x)/\sum_{x' \in \Omega} f(x')$.
Given a numerical database $\N$, the language of interval patterns $\mathcal{L}$ and $k \in \mathbb{N}$, the problem of sampling interval patterns {consists in} retrieving $k$ patterns $\IP_1,\ldots, \IP_k$ from $\mathcal{L}$, where each pattern $\IP_i$ is randomly drawn with replacement with a probability proportional to its value of $f$.
%desired measure's value.
Formally, this corresponds to:
\[ Sampling_{k}(\mathcal{L}, \N, f)= \bigcup_{i=1}^{k} ~\{\IP_i \in \mathcal{L}|~\IP_i \sim {f}(\IP_i) \}\]
This work focuses on the frequency measure, {i.e. $\freq(\IP_{i})$}, and demonstrates how the principle can be extended to the product of hyper-volume and frequency, {i.e. $\vol(\IP_{i}) \times \freq(\IP_{i})$}.
\section{Related Work}
\label{sec:rw}
We have organized the overview of the pattern sampling methods into three families: probabilistic, declarative and multi-step methods.
The first family is based on Markov chain Monte Carlo algorithms, where the target sampling distribution {corresponds} to the stationary distribution of a random walk. These methods can handle various pattern languages and interestingness measures. For example, {Hasan and Zaki} \cite{DBLP:journals/pvldb/alhassan2009} introduced the first pattern sampling approach that {handles a graph pattern language} and is biased towards a frequency measure. Similarly, Boley et al.~\cite{DBLP:conf/sdm/BoleyGG10} proposed sampling formal concepts using strictly positive interestingness measures, while Bendimerad et al.~\cite{DBLP:conf/ida/BendimeradLPRB20} focused on sampling tiles proportionally to a subjective interest. Camelin et al. \cite{DBLP:journals/dke/CamelinLPT25} define a compression-based approach leveraging the LCM algorithm \cite{DBLP:conf/fimi/UnoKA04} to sample frequent and closed itemsets. Finally, Opran et al.~\cite{RNTI/papers/1002998} introduced \textsc{SIMAS}, an interactive approach for sampling classification rules according to user-defined interest. Despite their accuracy, stochastic approaches often suffer from slow convergence.
The second family is based on declarative paradigms. \textsc{flexics}~\cite{DBLP:journals/datamine/DzyubaLR17} uses a SAT solver to sample itemsets and considers a wide range of measures. This approach randomly partitions the search space into different cells by recursively generating XOR constraints on variables associated with the items describing the patterns. However, modeling other pattern languages is challenging, as the encoding needs to be adapted for each language to ensure efficient implementation.
The third family is composed of multi-step methods. These methods partition the occurrences of the patterns into well-devised groups, enabling group sampling based on their weights, and then uniformly sampling a pattern within the selected group. The challenge is to find an appropriate multi-step decomposition to obtain the desired distribution. This approach was pioneered by Boley et al.~\cite{DBLP:conf/kdd/BoleyLPG11}, using a two-step draw to sample patterns according to the frequency. Diop et al.~\cite{DBLP:journals/kais/DiopDGLS20} extend this approach with a three-step method for sequential pattern sampling, using a frequency measure and a maximum length constraint. Soulet~\cite{DBLP:conf/f-egc/Soulet23} combines the sampling and constrained-based paradigms by designing an itemset pattern sampling approach enforcing a minimum frequency constraint. {Diop~\cite{DBLP:conf/pakdd/Diop22} proposed a two-step procedure to sample high average utility itemsets under length constraints. Finally, Diop and Plantevit~\cite{DBLP:conf/bigdataconf/DiopP24} extended this work to sampling high-utility patterns from knowledge graphs by converting them into quantitative transactional databases.}
Richer patterns such as subgroups are also addressed by pattern sampling. In this context, Moens and Boley~\cite{DBLP:conf/ida/MoensB14} present a weighted controlled pattern sampling method for exceptional model mining. This approach divides the search space and applies targeted weighting, balancing efficiency and relevance to discover exceptional patterns.
%\textcolor{red}{Furthermore, several works have exploited randomness during pattern space exploration. Egho et al.~\cite{DBLP:conf/icdm/EghoGBVC15} proposed \emph{MiSeRe}, an approach that uniformly samples sub-sequences for classification rule discovery. Sub-sequence candidates are uniformly drawn and rejected if their score is negative according to a user-defined \emph{level} measure. Bosc et al.~\cite{DBLP:journals/datamine/BoscBRK18} proposed an approach for subgroup discovery in numerical data. Subgroups (patterns) are extracted as conjunctions of inequalities through a Monte Carlo Tree Search (MCTS) algorithm that balances exploration and exploitation, ensuring good diversity among the extracted subgroups. Similarly, Belfodil et al.~\cite{DBLP:conf/pkdd/BelfodilBK18} introduced \emph{RefineAndMine}, a method for discovering optimal subgroups from numerical data. This approach uses an interval-based representation of patterns and relies on a sequence of progressively refined discretizations. It can incorporate randomness in the selection of cut-points to refine, improving the exploration of certain parts of the search space when multiple refinement options are available.
%These approaches all employ randomness to explore the search space and return a sample of patterns. In this sense, they can be viewed as sampling methods. However, unlike the output-based sampling approaches discussed earlier, they do not draw patterns according to a predefined target probability distribution. Instead, randomness is primarily used to diversify the discovered patterns, without theoretical guarantees on the sampling distribution.
%}
To the best of our knowledge, there is only one method for output space pattern sampling from numerical data~\cite{DBLP:conf/sdm/GiacomettiS18}. This method uses a multi-step approach. In contrast to \fips{} and \hips{}, it considers a continuous data space, samples according to a density measure, and requires the pattern size to be specified. {Our work falls into this class of approaches due to its ability to directly obtain the desired sampling distribution.}
%We have organized the overview of the pattern sampling methods into three families: probabilistic, declarative and multi-step methods.
%The first family is based on Markov chain Monte Carlo algorithms, where the target sampling distribution corresponds to the stationary distribution of a random walk. These methods can handle various pattern languages and interestingness measures. For example, Hasan and Zaki \cite{DBLP:journals/pvldb/alhassan2009} introduced the first pattern sampling approach that handles a graph pattern language and is biased towards a frequency measure. Similarly, Boley et al.~\cite{DBLP:conf/sdm/BoleyGG10} proposed sampling formal concepts using strictly positive interestingness measures, while Bendimerad et al.~\cite{DBLP:conf/ida/BendimeradLPRB20} focused on sampling tiles proportionally to a subjective interest. Camelin et al. \cite{DBLP:journals/dke/CamelinLPT25} define a compression-based approach leveraging the LCM algorithm \cite{DBLP:conf/fimi/UnoKA04} to sample frequent and closed itemsets. Finally, Opran et al.~\cite{RNTI/papers/1002998} introduced \textsc{SIMAS}, an interactive approach for sampling classification rules according to user-defined interest. Despite their accuracy, stochastic approaches often suffer from slow convergence.
%The second family is based on declarative paradigms. \textsc{flexics}~\cite{DBLP:journals/datamine/DzyubaLR17} uses a SAT solver to sample itemsets and considers a wide range of measures. This approach randomly partitions the search space into different cells by recursively generating XOR constraints on variables associated with the items describing the patterns. However, modeling other pattern languages is challenging, as the encoding needs to be adapted for each language to ensure efficient implementation.
%The third family is composed of multi-step methods. These methods partition the occurrences of the patterns into well-devised groups, enabling group sampling based on their weights, and then uniformly sampling a pattern within the selected group. The challenge is to find an appropriate multi-step decomposition to obtain the desired distribution. This approach was pioneered by Boley et al.~\cite{DBLP:conf/kdd/BoleyLPG11}, using a two-step draw to sample patterns according to the frequency. Diop et al.~\cite{DBLP:journals/kais/DiopDGLS20} extend this approach with a three-step method for sequential pattern sampling, using a frequency measure and a maximum length constraint. Soulet~\cite{DBLP:conf/f-egc/Soulet23} combines the sampling and constrained-based paradigms by designing an itemset pattern sampling approach enforcing a minimum frequency constraint.
%Richer patterns such as subgroups are also addressed by pattern sampling. In this context, Moens and Boley~\cite{DBLP:conf/ida/MoensB14} present a weighted controlled pattern sampling method for exceptional model mining. This approach divides the search space and applies targeted weighting, balancing efficiency and relevance to discover exceptional patterns.
%To the best of our knowledge, there is only one method for sampling patterns from numerical data~\cite{DBLP:conf/sdm/GiacomettiS18}. This method uses a multi-step approach. In contrast to \fips{} and \hips{}, it considers a continuous data space, samples according to a density measure, and requires the pattern size to be specified. Our work falls into this class of approaches due to its ability to directly obtain the desired sampling distribution.
\section{Sampling Interval Patterns Proportionally to their Frequencies}
\label{sec:contrib}
This section introduces \fips{}, an approach for sampling interval patterns proportionally to their frequency. The problem is formally defined as follows:
\[
Sampling_{k}(\mathcal{L}, \N, \freq) = \bigcup_{i=1}^{k} \{\IP_i \in \mathcal{L} \mid \IP_i \sim {\freq(\IP_i)}\}
\]
\subsection{\fips{} Key Ideas}
\label{sec:fips key ideas}
To draw a pattern proportionally to its frequency, it is necessary to determine the sum of the frequencies of all patterns in the solution space, i.e. $\sum_{\IP \in \mathcal{L}} \mathrm{freq}(\IP)$. Due to the size of the solution space $\lang$, it is intractable to compute this sum by enumerating each solution in the space. This limitation can be overcome by using a multi-step sampling procedure \cite{DBLP:conf/kdd/BoleyLPG11}.
The key idea is to compute weights by grouping patterns according to their occurrences in the set of objects. Each group is then drawn in proportion to its calculated weight, then a pattern is selected uniformly within the chosen group.
%
Determining the number of patterns covering an object in binary data is straightforward {as it is simply equal to $2^{|g|}$}, where $|g|$ is the number of items present in the object $g$. In contrast, for numerical data, this task is more complex. {Unlike itemsets, which are contained within each transaction, interval patterns are not necessarily contained within the objects. Moreover, computing the exact number of interval patterns covering an object $g$ requires, for each attribute, considering all possible intervals that include the values appearing in $g$.} To address this challenge, we define a function $\NIP{}$ (Number of Interval Patterns) that computes the exact number of interval patterns covering an object.
%Unlike itemsets, which are contained within each transaction, interval patterns are not necessarily included in the objects
%Determining the number of patterns covering an object in binary data is straightforward as it is simply equal to $2^{|g|}$, where $|g|$ is the number of items present in the object $g$. In contrast, for numerical data, this task is more complex, as it requires considering, for each attribute of an object, all the possible intervals of the other attributes covering the object. To address this problem, we define a function $\NIP{}$ (Number of Interval Patterns) that determines the exact number of interval patterns covering an object.
%
Using this function, it is possible to calculate the sum of the frequencies of all patterns (the normalisation constant $Z$, see Section~\ref{sec:algo echantillonnage}) by considering only the patterns which cover an object. Thus, this method computes a probability distribution proportional to frequency without enumerating the entire solution space.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\subsection{Counting the Number of Interval Patterns Covering an Object}
\label{sec:NIP}
This section defines the function $\NIP{} : \ig \rightarrow \mathbb{N}$ which counts the exact number of interval patterns covering the object $\ig \in \G$.
For a value $\val{}{} \in \N_\im$ appearing in an object $\ig \in \G$, we define:
\begin{itemize}
\item{ $\I{\val{}{}} = \{v \in \W_m \mid v \leq \val{}{}\}$ as the set of distinct values in attribute $\im$ that can be used as lower bounds in interval patterns covering the object $g$. }
\item {$\J{\val{}{}} = \{v \in \W_m \mid v \geq \val{}{}\}$ as the set of distinct values in attribute $\im$ that can be used as upper bounds in interval patterns covering the object $g$.}
\end{itemize}
%\begin{itemize}
% \item $\mathcal{U}(\val{}{}) = \{v \in \W_m \mid v < \val{}{}\}$, represents the set of distinct values of the attribute $\im$ forming lower bounds in the intervals containing the value $\val{}{}$.
% \item $\mathcal{A}(\val{}{}) = \{ v \in \W_m \mid v > \val{}{}\}$, represents the set of distinct values of the attribute $\im$ forming upper bounds in the intervals containing the value $\val{}{}$.
%\end{itemize}
The $\NIP(g)$ function is defined as follows:
%\begin{equation}
%\label{eq:NIP}
%\NIP(g) = \prod_{m \in \M} \quad |\mathcal{U}(\val{}{})|+ |\mathcal{A}(\val{}{})|+ |\mathcal{U}(\val{}{})|\cdot |\mathcal{A}(\val{}{})|+1
%\end{equation}
\begin{equation}
\label{eq:NIP}
\NIP(g) = \prod_{m \in \M} \quad {|\I{\val{}{}}| \cdot |\J{\val{}{}}|}
\end{equation}
{Equation~\ref{eq:NIP} relies on two terms. The first term $|\I{\val{}{}}|$ counts the number of distinct values in $\W_m$ that can serve as lower bounds for intervals associated to attribute $m$. Similarly, the second term $|\J{\val{}{}}|$ counts the number of distinct values in $\W_m$ that can serve as upper bounds for intervals associated to attribute $m$. The product $|\I{\val{}{}}| \cdot |\J{\val{}{}}|$ returns, for an object $g \in \G$ and an attribute $m \in \M$, the total number of intervals including the value $\val{}{}$. The exact number of interval patterns covering the object $\ig$ is given by the product of the possible intervals across all attributes in the database.}
%For each attribute $\im \in \M$, Eq.~\ref{eq:NIP} is defined over several terms, each representing the number of intervals that include the value $\val{}{}$. The term $|\mathcal{U}(\val{}{})|$ counts intervals where $\val{}{}$ is the upper bound, while $|\mathcal{A}(\val{}{})|$ counts the intervals where $\val{}{}$ is the lower bound. The product $|\mathcal{U}(\val{}{})| \cdot |\mathcal{A}(\val{}{})|$ represents intervals where $\val{}{}$ lies strictly between the possible bounds over the attribute $\im$. Finally, the unique interval where $\val{}{}$ is both the lower and the upper bound is added to the definition.
%The exact number of interval patterns covering the object $\ig$ is given by the product of the possible intervals across all attributes in the database.
\begin{example}
{
Consider the numerical database $\N$ presented in Table~\ref{dataset:nv:exemple}, the object $\ig_{3} \in \G$, and the attribute $m_{1} \in \M$. The distinct value of attribute $m_1$ in object $\ig_3$ is 3. The intervals associated to $m_1$ that include the value $3$ are the intervals $[2,3]$, $[2,4]$, $[2,6]$, $[3,3]$, $[3,4]$ and $[3,6]$.
%\begin{itemize}
% \item the value 3 is a lower bound~: $[3,4],~[3,6]$
% \item the value 3 is an upper bound~: $[2,3]$
% \item the value 3 is strictly included~: $[2,4],~[2,6]$
% \item the value 3 is both the lower and upper bound~: $[3,3]$
%\end{itemize}
Thus, the total number of intervals including the value 3 for the attribute $m_{1}$ is 6. Referring to the terms in Equation~\ref{eq:NIP}, we have: $|\I{3}|\cdot |\J{3}| = |\{2,3 \}| \cdot |\{3, 4, 6\}| =6$. The total number of interval patterns covering $\ig_3$ is obtained by multiplying the number of intervals for each attribute (see Equation~\ref{eq:NIP}): $6 \times 4 \times 5 = 120$.
}
\end{example}
\subsection{\fips{} Sampling Algorithm}
\label{sec:algo echantillonnage}
%This section describes the sampling algorithm \fips{} (see Algorithm~\ref{algo:FIPS}).
This section describes the \fips{} sampling procedure (see Algorithm~\ref{algo:FIPS}). The algorithm starts by determining the number of interval patterns covering each object in the database $\N$ (line~\ref{algo:0}). {This is performed by} using the $\NIP{}$ function (see Section~\ref{sec:NIP}). Line~\ref{algoFips:Step1} implements the first step of the two-step sampling procedure: an object $g$ is sampled with a probability proportional to the number of interval patterns covering $g$. Thus, the draw is biased to select patterns covering a high number of objects. Then, in the second step (line~\ref{algo:2}), an interval pattern covering at least the object $g$ is built attribute by attribute through successive interval samplings. Specifically, for each attribute $m \in \M$, two values {$a_m \in \I{\val{}{}}$ and $b_m \in \J{\val{}{}}$} are uniformly sampled within each respective set. {These values form the interval associated to attribute $m$ and are concatenated into the pattern under construction (line~\ref{algo:concat}). Once all attributes have been associated to an interval, the resulting interval pattern $\IP$ is returned (line~\ref{algo:returnIP}).} A sample of $k$ patterns is obtained by running $k$ times Algorithm~\ref{algo:FIPS}. %In the following section, we demonstrate that \fips{} samples patterns proportionally to their frequency.
\begin{algorithm}[h]
{
\textbf{Input :} A numerical database $\N$ \;
\textbf{Output :} An interval pattern $\IP$ sampled proportionally to its frequency \;
\nonl~\\
\textbf{Pre-processing:} Compute $w_{F}(g)= \NIP(g)$ for each object $g \in \G$\label{algo:0}\;
\nonl~\\
%\textbf{Step 1:} Draw an object $g \sim w(g)$ \label{algo:1}\;
\textbf{Step 1:} $g \sim w_{F}(g)$ \quad \tcc*{Draw an object g proportionally to its weight $w_{F}$} \label{algoFips:Step1}
%frequency.*/
\nonl~\\
\textbf{Step 2:} \label{algo:2}\
\tcc*{Draw uniformly an interval pattern $\IP$ covering $g$.}
$\IP \longleftarrow \langle \rangle$\;
\ForEach{attribute $m \in \M $}{
$a_m \longleftarrow$ Draw uniformly a value from {$\I{\val{}{}}$}\;
$b_m \longleftarrow $ Draw uniformly a value from {$\J{\val{}{}}$}\;
$\IP \longleftarrow \IP ++ [a_m,~ b_m] $ \tcc*{{Concatenate the interval $[a_m, b_m]$ to $\IP$ ($++$ denotes concatenation)}} \label{algo:concat}\
}
\nonl~\\
\textbf{Return } $\IP$ \label{algo:returnIP}\
}
\caption{{Sampling an interval pattern} proportionally to frequency (\fips{})
\label{algo:FIPS}}
\end{algorithm}
%\subsection{Key Theoretical Guarantees of \fips{} Algorithm}
%In this section, we present the theoretical foundations of our \fips\ approach, demonstrating that interval patterns are sampled in proportion to their frequencies. We then conduct a computational complexity analysis that considers both data pre-processing and the sampling procedure.
\subsection{Sampling Distribution}
In this section, we demonstrate that \fips{} samples interval patterns in proportion to their frequencies.
\vspace*{0.2cm}
\begin{property}
\label{prop:fips:distribution}
For a numerical database $\N$, Algorithm~\ref{algo:FIPS} samples an interval pattern $\IP$ proportionally to its frequency.
\end{property}
\begin{proof}
Let $\lang$ be the space of all interval patterns, and $Z = \sum_{\IP \in \lang } |\cover(\IP)|$ be the normalization constant, representing the sum of the frequencies of all interval patterns. Let $g^{*} \in \G$ be an object randomly drawn from step~1 of Algorithm~\ref{algo:FIPS}, and $\IP^{*}$ be an interval pattern sampled in step~2 of the same algorithm. Then:
\begin{align*}
P[\IP^{*}= \IP]&= \sum_{g ~\in~ \G} {P[\IP^{*}= \IP ~~\cap~~ g^{*}= g]}\\
&={\sum_{g ~\in~ \G} P[\IP^{*}= \IP ~|~ g^{*}= g] ~\cdot~P[g^{*}= g]}\\
&= \sum_{g ~\in~ {\cover(\IP)}} \frac{1}{NIP(g)} \frac{NIP(g)}{Z}\\
&= \sum_{\scriptsize g ~\in~{\cover(\IP)}} \frac{1}{Z} = \frac{|{\cover(\IP)}|}{Z} = \frac{{freq(\IP)}}{Z}
\end{align*}
where $Z= \sum_{g ~\in~ \G}~ NIP(g)$ (which is equal to $Z= \sum_{ \IP ~\in~ \mathcal{L}} ~|{\cover(\IP)}| $)
\end{proof}
%\begin{proof}
%Let $\lang$ be the space of all interval patterns, and $Z = \sum_{\IP \in \lang } |\cover(\IP$, $ \N)|$ be the normalization constant, representing the sum of the frequencies of all interval patterns. Let $g^{*} \in \G$ be an object randomly drawn from step~1 of Algorithm~\ref{algo:FIPS}, and $\IP^{*}$ be an interval pattern sampled in step~2 of the same algorithm. Then:
%\begin{align*}
%P[\IP^{*}= \IP]&= \sum_{g ~\in~ \G} P[\IP^{*}= \IP ~\land~ g^{*}= g]\\
%&= \sum_{g ~\in~ \cover(\IP)} \frac{1}{NIP(g)} \frac{NIP(g)}{Z}\\
%&= \sum_{\scriptsize g ~\in~\cover(\IP)} \frac{1}{Z} = \frac{|\cover(\IP)|}{Z} = \frac{freq(\IP)}{Z}
%\end{align*}
%where $Z= \sum_{g ~\in~ \G}~ NIP(g)$ (which is equal to $Z= \sum_{ \IP ~\in~ \mathcal{L}} ~|\cover(\IP)| $)
%\end{proof}
\subsection{Complexity Analysis of \fips{}}
In this section we conduct a computational complexity analysis for \fips{} that considers both data pre-processing and the sampling procedure.
\vspace*{0.2cm}
\begin{property}
The overall time complexity of the \fips{} approach for sampling a single interval pattern is:
\[
O( |\G| \cdot |\M| + {log|\G| + |\M|})
\]
\end{property}
\begin{proof}
The numerical database $\mathcal{N}$ contains $|\G|$ objects and $|\M|$ attributes. In the worst case, each attribute has $|\G|$ distinct values. The time complexity of the pre-processing step, which uses the \NIP{} function, is $O(|\G| \cdot |\M|)$. This corresponds to computing the candidate sets $\mathcal{U}(\val{}{})$ and $\mathcal{A}(\val{}{})$ for each of the $|\M|$ attributes, where each computation takes $O(|\G|)$ time.
The time complexity of sampling an interval pattern using Algorithm~\ref{algo:FIPS} combines the cost of drawing an object in step~1 and sampling intervals for each attribute in step~2. {Drawing an object proportionally to its $\NIP(g)$ value is done by computing the cumulative sum of all $\NIP$ values, generating a random number scaled by the total sum, and locating the corresponding object via dichotomic search in $O(\log |\G|)$. Sampling intervals uniformly in step~2 takes $O(|\M|)$. Therefore, the overall asymptotic time complexity, including both the pre-processing and sampling phases, is $O(|\G| \cdot |\M| + \log |\G| + |\M|)$.}
%The time complexity of sampling an interval pattern using Algorithm~\ref{algo:FIPS} is the sum of the complexities of the two steps: the dichotomic search in Step~1 takes $O(\log|\G|)$ time, and the uniform interval sampling in Step~2 takes $O(|\M|)$ time. Therefore, the overall asymptotic time complexity, including both pre-processing and sampling phases, is $O(|\G| \cdot |\M| + \log|\G| + |\M|)$.
\end{proof}
Sampling $k$ interval patterns requires $O(k(\log|\G| + |\M|))$ time in addition to the one-time pre-processing cost of $O(|\G| \cdot |\M|)$.
\input{tex/hyperVolume}
\vspace*{0.2cm}
In an interactive pattern mining process, the analyst is often interested in patterns that cover a substantial region of the search space. For interval patterns, this corresponds to patterns with a large hyper-volume. However, considering only hyper-volume may lead to patterns covering only few observations.
%However, considering patterns with large hyper-volumes without any other control may lead to patterns covering only few observations.
To avoid such undesirable patterns, we propose to consider both hyper-volume and frequency. The next section presents \hips{}, an extension of \fips{}, which samples patterns proportionally to the product of their frequency and hyper-volume.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%\section{\hfips{}: extending \fips{} to hyper-volume}
\section{\hips{}: Incorporating Hyper-volume to \fips{}
%\hips{}: add hyper-volume to \fips{}
}
\label{contrib:HFIPS}
% \todo[inline]{DB: J'ai gardé un backup du travail deja fait par lamine dans le fichier Backup hips}
% AO
%This section presents \hips{}. \hips{} draws patterns proportionally to the product of their frequency and hyper-volume. The problem is formally defined as follows:
This section presents \hips{}, a method that samples patterns proportionally to the product of their frequency and hyper-volume. The problem is formally defined as follows:
\[
Sampling_{k}(\mathcal{L}, \N, \freq \times \vol) = \bigcup_{i=1}^{k} \{\IP_i \in \mathcal{L} \mid \IP_i \sim {\freq(\IP_i) \cdot \vol(\IP_i)}\}
\]
\subsection{\hips{} Key Ideas}
\hips{} is built on the same core principles as \fips{} (cf. Section~\ref{sec:fips key ideas}). It computes, for each pattern in the solution space, the product of its hyper-volume and frequency without explicitly enumerating the entire space. This allows \hips{} to obtain the total sum $\sum_{\IP \in \mathcal{L}} (\freq(\IP) \cdot \vol(\IP))$ while avoiding a full enumeration. To achieve this, we introduce a new function, denoted $IPH$, which takes as input an object $g \in \G$ and computes the sum of the hyper-volumes of the patterns covering $g$. Summing the $IPH$ values across all objects leads to the normalization constant $Z$, as each pattern contributes proportionally to its frequency (see Section~\ref{sec:freqandvolume}).
%Version Bruno
%\hips{} is built on the same core principles as \fips{} (cf. section~\ref{sec:fips key ideas}). \hips{} requires to compute for each pattern of the solution space the product of its hyper-volume by its frequency and then the sum of the results of all these multiplications, i.e., $\sum_{\IP \in \mathcal{L}} (\freq(\IP) \times \vol(\IP))$,
%while avoiding a full enumeration of the pattern space.
%In line with the multi-step approaches such as \cite{DBLP:conf/kdd/BoleyLPG11},
%\hips{} builds upon the same core principles as \fips{}.
%As part of the multi-step approaches \cite{DBLP:conf/kdd/BoleyLPG11}, \hips{} relies on the same key ideas as \fips{}.
%To bias the sampling process toward more frequent patterns with wider intervals across all attributes,
%\hips{} requires determining the sum of all hyper-volumes times the frequencies of all patterns in the solution space, i.e., $\sum_{\IP \in \mathcal{L}} (\freq(\IP) \times \vol(\IP))$,
%while avoiding exhaustive enumeration
%while avoiding full enumeration of the pattern space.
%
%To achieve this goal, we introduce a new function, called $IPH$, which takes as input an object $g\in\G$ and computes the sum of the hyper-volumes of the patterns covering the object $g$.
%To achieve this, we define a new function $IPH$ which, for each object $g$ in the database, computes the sum of the hyper-volumes of the patterns covering $g$.
%
%For that purpose, we introduce a new function, called $IPH$, which takes an object $g\in\G$ as input and computes the sum of the hyper-volumes of the patterns covering the object $g$. Summing the $IPH$ values across all objects ensures that each pattern contributes proportionally to its frequency, thus directly producing the normalization constant $Z$ (see Section~\ref{sec:freqandvolume}). %This allows an efficient computation of a probability distribution proportional to the hyper-volume and the frequency.
%Since patterns cover multiple objects, these patterns are taken into consideration as many times as their frequencies. Thus, summing $IPH$ over all objects directly leads to the normalisation constant $Z$ (see section~\ref{sec:freqandvolume}) without requiring an exhaustive exploration of the solution space. This allows an efficient computation of a probability distribution proportional to the hyper-volume times the frequency.
%Since patterns cover multiple objects, they are counted as many times as their frequencies. Thus, applying this function to all objects in the database will result in the sum of the hyper-volumes times the frequency of all interval patterns in the solution space.
%
\subsection{Efficiently Counting Total Hyper-volumes }
\label{sec:IH&TIA}
This section defines the function $IPH : \ig \rightarrow \mathbb{N}$, which computes the sum of the hyper-volumes of interval patterns covering the object $\ig \in \G$. $IPH$ is based on the idea of summing the lengths of all intervals that contain an object. We start by introducing this notion.
\vspace{0.5cm}
\textbf{Total Intervals Lengths.} To efficiently compute the hyper-volumes of interval patterns covering an object $\ig \in \G$, we first consider, for each attribute $m \in \M$, the sum of the lengths of all intervals containing the value $\val{}{} \in \N_\im$. To this end, we use the sets $\I{\val{}{}}$ and $\J{\val{}{}}$ introduced previously:
\begin{itemize}
\item $\I{\val{}{}} = \{v \in \W_m \mid v \leq \val{}{}\}$ as the set of distinct values in attribute $\im$ that can be used as lower bounds in interval patterns covering the object $g$.
\item $\J{\val{}{}} = \{v \in \W_m \mid v \geq \val{}{}\}$ as the set of distinct values in attribute $\im$ that can be used as upper bounds in interval patterns covering the object $g$.
\end{itemize}
%\begin{itemize}
% \item $\I{\val{}{}} = \{v \in \W_m \mid v \leq \val{}{}\}$ be the set of distinct values appearing in attribute $\im$ that can be used as lower bounds in interval patterns covering the object $g$.
% \item $\J{\val{}{}} = \{v \in \W_m \mid v \geq \val{}{}\}$ be the set of distinct values appearing in attribute $\im$ that can serve as upper bounds in interval patterns covering the object $g$.
%\end{itemize}
The $TIL$ (Total Interval Lengths) function can be defined as follows:
%This is done using the $TIL$ function (Total Interval Lengths):
\begin{equation}
\label{eq:TIL}
\TIL{m}{g} = \left(\sum_{x \in \J{\val{}{}}} x \cdot |\I{\val{}{}} | \right) - \left(\sum_{x \in \I{\val{}{}}} x \cdot |\J{\val{}{}} |\right)
\end{equation}
where, the first term $(\sum_{x \in \J{\val{}{}}} x \cdot |\I{\val{}{}} |)$ represents
the sum of interval lengths by summing directly all the values of the upper bounds in $\J{\val{}{}}$ and
multiplying this sum by the number of lower bounds in $\I{\val{}{}}$ since all upper bounds will be combined with each lower bound to form the intervals on the attribute including $\val{}{}$.
%the total contribution of upper bounds to the total sum of interval lengths.
%Specifically, it sums all upper bounds in $\J{\val{}{}}$ and multiplies the result by the number of lower bounds in $\I{\val{}{}}$, since each upper bound is combined with each lower bound to form a valid interval.
However, this term alone overestimates the total length of all intervals, since it includes the upper bounds without subtracting the corresponding lower bounds that define the actual interval lengths.
On the opposite side, the second term $(\sum_{x \in \I{\val{}{}}} x \cdot |\J{\val{}{}} |)$, corrects this overestimation by subtracting
the sum of interval lengths by summing directly all the values of the lower bounds in $\I{\val{}{}}$ and
multiplying this sum by the number of upper bounds in $\J{\val{}{}}$. Thus, ensuring with the first term that the length of each interval is computed exactly once. Consequently, the $TIL$ function leads to the exact sum of the lengths of all intervals containing $\val{}{}$ over the attribute $m$.
%the total contribution of the lower bounds to the same sum (C'est faux ce n'est pas la même somme ). It multiplies the sum of the lower bounds by the number of upper bounds, thus ensuring that the length of each interval is computed exactly once. Combined, these two terms lead to the exact sum of the lengths of all intervals containing $\val{}{}$ on attribute $m$.
\begin{example}
Consider the numerical dataset $\N$ shown in Table~\ref{dataset:nv:exemple}. For simplicity, we focus on attributes $m_1$ and $m_2$. The Total Intervals Length for the object $g_1$ is computed for both attributes as follows:
\begin{itemize}
\item Total interval lengths for $m_1$: The value of $g_1$ for $m_1$ is $2$. $\TIL{m_1}{g_1}$ computes the total length of all intervals that contain the value $2$.
Here, $\I{2} = \{2\}$ and $\J{2} = \{2,3,4,6\}$. The first part, $|\I{2}| \cdot \sum \J{2} = 1 \cdot (2+3+4+6)$, sums all possible upper bounds for intervals having $2$ as a lower bound, assuming they can all be paired with this single lower bound. However, this overestimates the total length, as it does not take into account the lower bound value itself. To correct this, the second term $|\J{2}| \cdot \sum \I{2} = 4 \cdot 2$ subtracts the repetitive contribution of the lower bound $2$, once for each possible upper bound. Therefore:
\begin{align*}
\TIL{m_1}{g_1} &= |\{2\}| \times (2+3+4+6) - |\{2,3,4,6\}| \times 2 \\
&= (1 \times 15) - (4 \times 2) = 7
\end{align*}
This leads to the exact sum of lengths of the intervals $[2,2]$, $[2,3]$, $[2,4]$ and $[2,6]$, i.e., $0 + 1 + 2+ 4 = 7$.
\item Total interval lengths for $m_2$: The value of $g_1$ for $m_2$ is $8$. Similarly, $\TIL{m_2}{g_1}$ computes the total length of all intervals that contain the value $8$. Here, $\I{8} = \{7,8\}$ and $\J{8} = \{8,9,12\}$. The term $|\I{8}| \cdot \sum \J{8} = 2 \cdot (8+9+12)$ sums all upper bounds for each compatible lower bound. This overestimates the total interval lengths. The correction term $|\J{8}| \cdot \sum \I{8} = 3 \cdot (7+8)$ subtracts the repetitive contribution of the lower bounds. Therefore:
\begin{align*}
\TIL{m_2}{g_1} &= |\{7,8\}| \times (8+9+12) - |\{8,9,12\}| \times (7+8)\\
&= (2 \times 29) - (3 \times 15) = 13
\end{align*}
The resulting value corresponds to the exact sum of lengths of the intervals $[7,8]$, $[7,9]$, $[7,12]$ $[8,8]$,$[8,9]$ and $[8,12]$ i.e., $1 + 2 + 5+ 0 + 1+ 4 = 13$.
\end{itemize}
\end{example}
\textbf{Total Hyper-volumes.} Once the total lengths of the intervals
have been computed for each attribute $m \in \M$, the product of these interval lengths across all attributes gives the total hyper-volume of the interval patterns covering the object $g$. which is formally defined as follows:
\begin{equation}
\label{eq:IPH}
\IH{g} = \prod_{m \in \M} \TIL{m}{g}
\end{equation}
%
The proof of the $IPH$ function is provided in \ref{appendix:tilProof}
%The hyper-volume of an interval pattern is defined as the product of the lengths of all intervals composing the pattern (see Section~\ref{sec:context}). To compute the sum of the hyper-volumes of interval patterns covering $g$, one approach is to determine, for each attribute $m \in \M$, the sum of the lengths of all intervals containing $\val{}{}$. In equation \ref{eq:IPH}, this sum of lengths correspond to the terms between parentheses. Then, the product of these sums over all attributes lead to the total sum of the hyper-volumes of all interval patterns covering $g$.
{
%\color{purple}
%\todo[inline]{AO : inclure g5 dans tes calculs, tu peux ne garder qu'1 ou 2 attributs mais faut que les calculs suivent tous %les objets de la table 1. Tu risques de dérouter le relecteur. Je saute cette partie le temps de la retravailler.
%{\textbf{DB: Corrigé}}
}
\begin{example}
Continuing from the previous example, the sum of the hyper-volumes of all interval patterns covering object $g_1$ is obtained by multiplying the total interval lengths for attributes $m_1$ and $m_2$:
\begin{align*}
\IH{g_1} &= \TIL{m_1}{g_1} {~\cdot~} \TIL{m_2}{g_1} \\
&= 7 \times 13 = 91
\end{align*}
\end{example}
\subsection{\hips{} Sampling Algorithm}
This section describes the \hips{} sampling algorithm (see Algorithm~\ref{algo:hv_sampling}). The algorithm starts by computing, for each object $g$ in the database $\N$, the total hyper-volume of interval patterns covering $g$ (line~\ref{algohips:preprocessing}), using the \texttt{IPH} function (see Section~\ref{sec:IH&TIA}). Line~\ref{algoFips:Step1} implements the first step of the two-step sampling procedure, where an object $g$ is drawn with probability proportional to {the sum of hyper-volumes of the interval patterns that cover it}. In the second step (from line~\ref{algohips:step2}, to line \ref{algohips:concactinterval}), an interval pattern covering at least the selected object $g$ is sampled. The pattern is constructed attribute by attribute. For each attribute $m \in \M$, a lower bound $a_m$ is drawn from the set of candidate distinct values $\I{\val{}{}}$, using a weight function $w_{lb}$ that incorporates the cumulative contribution of potential upper bounds (line~\ref{algohips:step2lb}). This bias favors smaller values of $a_m$, which are more likely to lead to larger intervals. Then, an upper bound $b_m$ is sampled from the set of candidate values $\J{\val{}{}}$, using a weight function $w_{ub}$ that promotes wider intervals by taking into account the difference $b_m - a_m$ (line~\ref{algohips:step2ub}). Once both bounds are drawn, the corresponding interval is added to the {pattern under construction} (line~\ref{algohips:concactinterval}).
\begin{algorithm}[h!]
\caption{{Sampling an interval pattern proportionally to the product of its hyper-volume and frequency} (\hips{}) }
\label{algo:hv_sampling}
\textbf{Input:} A numerical database $\N $\;
\textbf{Output:} An interval pattern $\IP$ sampled proportionally to its hyper-volume $\times$ frequency\label{algo1:preprocessing}\;
\bigskip
\textbf{Pre-processing:} Compute $w_{HF}(g)= \IH{g}$ for each object $g \in \G$ \label{algohips:preprocessing}\;
\bigskip
\textbf{Step 1:} Draw an object $g \in \G$ proportionally to it weight $w_{HF}$\label{algohips:step1} \;
\bigskip
$\IP \longleftarrow \langle\rangle $ \tcc*{ Interval pattern under construction}
\textbf{Step 2:}
\ForEach{attribute $m \in \M $}{ \label{algohips:step2}
\bigskip
$\I{\val{}{}} = \{v \in \W_m \mid v \leq \val{}{}\}$ \tcc*{ Compute the candidate lower bound set}
\bigskip
$\J{\val{}{}} = \{v \in \W_m \mid v \geq \val{}{}\}$ \tcc*{ Compute the candidate upper bound set}
\bigskip
Draw a lower bound $a_m \in \I{\val{}{}}$ proportionally to $w_{lb}$, where: \label{algohips:step2lb}
\begin{algomathdisplay}
w_{lb}(a_m) = \left(\sum_{x \in J_{g,m}} x\right) - |J_{g,m}| \cdot a_m
\end{algomathdisplay}
%\[
%w_{lb}(a_m) = \left(\sum_{x \in J_{g,m}} x\right) - |J_{g,m}| \cdot a_m
%\]%\;
Draw an upper bound $b_m \in \J{\val{}{}}$ proportionally to $w_{ub}$, where: \label{algohips:step2ub}
\begin{algomathdisplay}
w_{ub}(b_m) = b_m - a_m
\end{algomathdisplay}
%\[
%w_{ub}(b) = b_m - a_m
%\]%\;
$\IP \longleftarrow \IP ++ [a_m, b_m]$ \label{algohips:concactinterval} \tcp*{ Adding $[a_m, b_m]$ to the interval pattern in construction}
}
\Return{$\IP$} %= \langle [a_m, b_m] \rangle_{m \in \{1, \dots, |\M|\}}
\end{algorithm}
% \subsection{Key Theoretical Guarantees of \hips{} Algorithm}
%\label{sec:theoretical_analysis}
% This section depicts the theoretical foundations of our \hips\ approach. We demonstrate that interval patterns are sampled in proportion to the product of their hyper-volumes and frequencies. We then conduct a computational complexity analysis that considers {both the data pre-processing} and the sampling procedure.
\subsection{Sampling Distribution}
\label{sec:hips_Sampling Distribution}
In this section, we demonstrate that \hips{} samples interval patterns in proportion to their frequencies $\times$ hyper-volumes.
\begin{property}
For a numerical database $\N$, the \hips\ algorithm samples an interval pattern $\IP$
%= \langle [\inff{w_m}, \supp{w_m}] \rangle_{m \in \{1, \ldots, |\M|\}}$
with probability proportional to the product of its hyper-volume and frequency. Formally:
\[
P[\IP] \propto \vol(\IP) \cdot \freq(\IP)
\]
\end{property}
\begin{proof}
Let $\mathcal{L}$ be the set of all interval patterns. We define the normalization constant $Z$ as:
\[
Z = \sum_{\IP \in \mathcal{L}} \vol(\IP) \cdot \freq(\IP) = \sum_{g ~\in~ \G}~ \IH{g}.
\]
%
The \hips{} sampling procedure operates in two sequential steps.
%\noindent \textbf{Object Selection:}
First, an object $g \in \G$ is sampled with a probability
%\[
$P(g) = \frac{\IH{g}}{\sum_{g' \in \G} \IH{g'}}.$
%\]
% where $\IH{g}$ denotes the Interval hyper-volume of $g$.
%\noindent \textbf{Interval Construction:}
Second, given $g$ from the first step, an interval pattern $\IP$ that covers $g$ is drawn with a conditional probability proportional to its weight.
We demonstrate first that
%\todo[inline]{DB: Pas sur du $w(\IP) $. Peut etre le remplacer par $\vol(\IP)$?}
%\[
%$P[\IP \mid g] = \frac{w(\IP)}{\IH{g}}.$
%AO: %$P[\IP \mid g] = \frac{\prod_{[\inff{w_m},\supp{w_m}]\in \IP} (\supp{w_m}-\inff{w_m})}{\IH{g}}.$
$P[\IP \mid g] = \frac{\vol(\IP)}{\IH{g}}.$
%\]
Then, we demonstrate that the overall probability of drawing a particular interval pattern $\IP$ by \hips{} is equal to $P[\IP] = \frac{\vol(\IP)}{Z} \times \freq(\IP)$.
For a given object $g \in \G$ and for each attribute $m \in \M$, \hips{} constructs an interval $[\inff{w_m}, \supp{w_m}]$ by performing two sequential sampling steps. Consider the candidate lower and upper bound sets: $\I{\val{}{}} = \{v \in \W_m \mid v \leq \val{}{}\} $ and $ \J{\val{}{}} = \{v \in \W_m \mid v \geq \val{}{}\}.$
\begin{enumerate}
\item \textbf{Lower Bound Selection:}
Each candidate lower bound $v \in \I{\val{}{}}$ is assigned a weight $w_{lb}(v)$. The probability of selecting a specific lower bound $\inff{w_m}$ is given by:
%\[
%P(\inff{w_m} \mid g) = \frac{w(\inff{w_m})}{\sum_{d \in \I{\val{}{}}} w(d)} \text{ with } w(d) = \left(\sum_{x \in \J{\val{}{}}} x\right) - |\J{\val{}{}}| \cdot d.
%\]
\begin{align*}
P(\inff{w_m} \mid g)
&= \frac{w_{lb}(\inff{w_m})}{\sum_{v \in \I{\val{}{}}} w_{lb}(v)}
\quad \text{with} \\
w_{lb}(v)
&= \left(\sum_{x \in \J{\val{}{}}} x\right) - |\J{\val{}{}}| \cdot v
\end{align*}
The normalizing constant $\sum_{v \in \I{\val{}{}}} w_{lb}(v)$ corresponds to the Total Intervals Length (cf. Equation \ref{eq:TIL}). This allows the normalization constant to be computed directly as:
\begin{align*}
\sum_{v \in \I{\val{}{}}} w_{lb}(v) &= |\I{\val{}{}}| \cdot \left(\sum_{x \in \J{\val{}{}}} x\right) - |\J{\val{}{}}| \cdot \left(\sum_{v \in \I{\val{}{}}} v\right)\\
&= \TIL{m}{g}.
\end{align*}
Then: \[P(\inff{w_m} \mid g) = \frac{w_{lb}(\inff{w_m})}{\TIL{m}{g}}\]
\vspace*{0.2cm}
\item \textbf{Upper Bound Selection:}
Once the lower bound $\inff{w_m}$ has been sampled, the upper bound $\supp{w_m}$ is selected from the set $\J{\val{}{}}$ according to the following probability:
\[
P(\supp{w_m} \mid g, \inff{w_m}) = \frac{\supp{w_m} - \inff{w_m}}{w_{lb}(\inff{w_m})}.
\]
This formulation favors larger intervals, as the probability increases with the difference $\supp{w_m} - \inff{w_m}$.
\end{enumerate}
Therefore, the overall conditional probability of selecting the interval $[\inff{w_m}, \supp{w_m}]$ for an attribute $m$ is given by:
\begin{align*}
P\bigl([\inff{w_m}, \supp{w_m}] \mid g\bigr)
&= P(\inff{w_m} \mid g) \cdot P(\supp{w_m} \mid g, \inff{w_m}) \\
&= \frac{w_{lb}(\inff{w_m})}{\TIL{m}{g}} \cdot \frac{\supp{w_m} - \inff{w_m}}{w_{lb}(\inff{w_m})} \\
&= \frac{\supp{w_m} - \inff{w_m}}{\TIL{m}{g}}.
\end{align*}
Since the intervals for different attributes are selected independently, the probability of constructing the entire interval pattern $\IP = \langle [\inff{w_m}, \supp{w_m}] \rangle_{m \in \{1, \dots, |\M|\}}$ covering an object $g \in \G$ is:
\begin{align*}
%P[\IP \mid g] &= \prod_{m \in \M} \frac{\supp{w_m} - \inff{w_m}}{\TIL{m}{g}}
P[\IP \mid g] &= \frac{\prod_{[\inff{w_m},\supp{w_m}] \in \IP} (\supp{w_m} - \inff{w_m})}{\prod_{m \in \M} \TIL{m}{g}}
\end{align*}
%When decomposed, this corresponds to:
Using the definition of hyper-volume and Equation~\ref{eq:IPH}, we have:
\begin{align*}
%P[\IP \mid g] &= \frac{\prod_{m \in \M} (\supp{w_m} - \inff{w_m})}{\prod_{m \in \M} \TIL{m}{g}} = \frac{\vol(\IP)}{\IH{g}}
%AO:%P[\IP \mid g] &= \frac{\prod_{[\inff{w_m},\supp{w_m}] \in \IP} (\supp{w_m} - \inff{w_m})}{\IH{g}}
P[\IP \mid g] &= \frac{\vol(\IP)}{\IH{g}}
\end{align*}
By the law of total probability \cite{feller1991introduction},
%\todo{BC : c'est connu/évide la signification "the law of total probability ? Référence ? LD: c'est une évidence je pense mais j'ai ajouté la ref},
the overall probability of drawing a particular interval pattern $\IP$ is
\[
P[\IP] = \sum_{g \in \cover(\IP)} P[\IP \mid g] \cdot P(g).
\]
Where $P(g)$ is the probability of sampling $g$ according to Step 1 of Algorithm~\ref{algo:hv_sampling}.
Substituting the expressions from the two steps results in:
\[
P[\IP] = \sum_{g \in \cover(\IP)} \frac{\vol(\IP)}{\IH{g}} \cdot \frac{\IH{g}}{\sum_{g' \in \G} \IH{g'}}
= \frac{\vol(\IP) \cdot \freq(\IP)}{\sum_{g' \in \G} \IH{g'}}.
\]
Defining $Z = \sum_{g' \in \G} \IH{g'}$, the normalization constant, completes the derivation:
\[
P[\IP] = \frac{\vol(\IP)}{Z} \times \freq(\IP).
\]
%Hence the result.
\end{proof}
\subsection{Complexity analysis of \hips{}}
We now analyze the time complexity of the \hips{} approach. The analysis is structured into three stages: data pre-processing, object sampling and interval construction.
\begin{property}
The overall time complexity of \hips{} for sampling one interval pattern is:
%For each attribute $m \in \M$, let $k_m = |\mathcal{D}_m|$ denote the number of distinct values it contains, and let $k = \max_{m \in \M} k_m$. Assuming $k_m \ll n$ for all $m \in \M$,
\[
O\Big(|\G| \cdot |\M|+ log(|\G|)+ |\M|\cdot (|\G| + \log |\G|) \Big).
\]
\end{property}
%\todo[inline]{LD: Cette complexité est énorme, et j'espère que tu n'as pas implémenté ta solution de cette manière. Mais cela peut expliquer le temps CPU sur tes XPs. En effet, dans l'algo HFIPS, les lignes 7, 8 et 9 doivent être exécutées une seule fois pendant la phase de pondération, car ces poids, pour un objet et un attribut donnés, ne changent pas (la valeur reste constante). Seule la valeur obtenue à la ligne 10 est inconnue, car elle dépend de la borne inférieure tirée. En réalité, on devrait extraire un objet en temps logarithmique, soit $O(\log(|G|))$, une borne inférieure également en $O(\log(|G|))$ (en supposant que toutes les valeurs de l'attribut sont distinctes, sinon $O(\log(|k|))$ avec $k < |G|$) et enfin, une borne supérieure en $O(|G|+\log(|G|))$ [j'avais oublié $|G|$ dans ma première formulation]. Ce qui donne au final une complexité de $O(3\log(|G|)+|G|)$. Ainsi, pour $K$ motifs sur une dimension comportant $m$ attributs, la complexité globale devrait être $O(K \cdot m \cdot (\log(|G|)+|G|))$.}
%\begin{property}
%%[Overall Complexity]
%Let $n = |\G|$ be the number of objects in the database, $l = |\M|$ the number of attributes, and $K$ the desired sample size. For each attribute $m \in \M$, let $k_m = |\mathcal{D}_m|$ denote the number of distinct values it contains, and let $k = \max_{m \in \M} k_m$. Assuming $k_m \ll n$ for all $m \in \M$, the overall time complexity of the \hips{} sampling approach is:
%\[
%O\Big(l \cdot n \log n + K \cdot l \log k\Big).
%\]
%\end{property}
\begin{proof}
The numerical database $\mathcal{N}$ contains $|\G|$ objects and $|\M|$ attributes. In the worst case, each attribute has $|\G|$ distinct values. For every object $g$ and attribute $m$, the candidate sets $\I{\val{}{}}$ and $\J{\val{}{}}$ are computed using a linear scan over $\W_m$, requiring $O(2 \cdot |\G|)$ time.
The overall sampling procedure can be divided into three phases:
\begin{itemize}
\item \textbf{Pre-processing:} For each object $g \in \G$, we compute its weight using the function $\IH{g}$. This involves computing the function $\TIL{m}{g}$, for every attribute $m \in \M$, which takes $O(2 \cdot |\G|)$ time per attribute. Since $\IH{g}$ is the product over all attributes, the total pre-processing time is $O(|\M| \cdot 2 \cdot |\G|)$.
\item \textbf{Step 1:} An object $g \in \G$ is drawn proportionally to its weight $\IH{g}$. This step is performed using a dichotomic search, which has a time complexity of $O(\log |\G|)$.
\item \textbf{Step 2:} For each attribute $m \in \M$, the following substeps are performed:
\begin{enumerate}
% \item Compute the sets $\I{\val{}{}}$ and $\J{\val{}{}}$ for the sampled object $g$, in $O(2 \cdot |\G|)$ time. \todo[inline]{LD: $\I{\val{}{}}$ et $\J{\val{}{}}$ ont été normalement déjà calculés pendant le prétraitement. Pourquoi vouloir les recalculer? Si tu le fais pour chaque attribut ça devient énorme et je pense que le mieux c'est de les garder en mémoire. On doit y accéder directement en O(1).}
\item For each lower bound candidate $a_m \in \I{\val{}{}}$, compute its weight using $w_{lb}$ in $O(|\G|)$ time.
\item Draw a lower bound $a_m$ proportionally to $w_{lb}$ using a dichotomic search in $O(\log |\G|)$.
\item For each upper bound candidate $b_m \in \J{\val{}{}}$, compute its weight using $w_{ub}$ in $O(|\G|)$ time.
\item Draw an upper bound $b_m$ proportionally to $w_{ub}$ using a dichotomic search in $O(\log |\G|)$.
\end{enumerate}
Hence, Step 2 has an asymptotic total complexity of
\[
O\big(|\M| \cdot (|\G| + \log |\G|)\big).
\]
\end{itemize}
By combining all three phases, the asymptotic overall time complexity of the \hips{} sampling procedure is:
\[
O\Big(|\G| \cdot |\M|+ log(|\G|)+ |\M|\cdot (|\G| + \log |\G|) \Big).
\]
%\noindent This concludes the proof.
\end{proof}
In practice, an analyst may wish to sample k patterns. Repeat \hips{} $k$ times requires $O(k \cdot (log(|\G|) + |\M| \cdot (|\G| + log(|\G|))))$ in addition to the pre-processing time which is in O($|\M| \cdot |\G|$). However, this complexity can be reduced by storing the sets $\I{\val{}{}}$ and $\J{\val{}{}}$ in memory during pre-processing, which avoids recomputing them during sampling.
% Thus, sampling $k$ interval patterns requires $O(k \cdot (log(|\G|) + |\M| \cdot (|\G| + log(|\G|))))$ in addition to the pre-processing time which is in O($|\M| \cdot |\G|$). This complexity can be reduced by storing the sets $\I{\val{}{}}$ and $\J{\val{}{}}$ in memory during pre-processing, which avoids recomputing them during sampling.
%\noindent This theoretical analysis reveals that the computational cost is primarily driven by the data pre-processing stage, where sorting and extracting distinct values across $m$ attributes incurs a cost of $O(m \cdot n \log n)$. The sampling phase, although influenced by the number of patterns $K$ and the maximum number $k$ of distinct values per attribute, remains efficient with a complexity of $O(K \cdot m \log k)$. In practice, since $k$ is typically much smaller than $n$, the overall approach scales well even for large database.
\subsection{{Sampling distribution and IPH}}
% Object Sampling Probability and Hyper-Volume Bias in IPH
{The following property quantifies the sampling probability of an object to the hyper-volumes of its interval patterns. We show that an object $g$ is more likely to be sampled than under uniform sampling if and only if the sum of the hyper-volumes of the interval patterns that cover $g$ exceeds the average hyper-volume.}
%{(higher total hyper-volume coverage)}
% The density awareness property links the sampling probability of an object to the hyper-volumes of its interval patterns. We now show that an object $g$ is more likely to be sampled than under uniform sampling if and only if the sum of hyper-volumes of the interval patterns that cover it exceeds the average. It ensures that objects in denser regions of the attribute space are more likely to be sampled. %{(higher total hyper-volume coverage)}
\begin{property}
The sampling distribution satisfies:
\[
P(g) \geq \frac{1}{|\G|} \iff \IH{g} \geq \frac{1}{|\G|} \sum_{g' \in \G} \IH{g'}%,
\]
% which prioritizes objects in dense regions through hyper-volume weighted sampling
\end{property}
\begin{proof}
By construction, our algorithm selects an object $g \in \G$ with probability
\[
P(g) = \frac{\IH{g}}{\sum_{g' \in \G} \IH{g'}}.
\]
An object $g$ is therefore sampled with a probability at least as high as under uniform sampling (i.e., $\frac{1}{|\G|}$) if and only if its hyper-volume $\IH{g}$ is at least equal to the average hyper-volume:
\[
P(g) \geq \frac{1}{|\G|} \iff \IH{g} \geq \frac{1}{|\G|} \sum_{g' \in \G} \IH{g'}.
\]
Let $A = \sum_{g' \in \G} \IH{g'}$ denote the total hyper-volume. Then:
\[
P(g) = \frac{\IH{g}}{A}.
\]
The inequality $P(g) \geq \frac{1}{|\G|}$ is equivalent to:
\[
\frac{\IH{g}}{A} \geq \frac{1}{|\G|} \quad \Longleftrightarrow \quad \IH{g} \geq \frac{A}{|\G|}.
\]
Since $\frac{A}{|\G|}$ corresponds to the average hyper-volume, this shows that $g$ is more likely to be sampled than under uniform sampling if and only if its hyper-volume exceeds the average.
%This concludes the proof.
\end{proof}
%}
\section{Experiments and Results}
\label{sec:exp}
This section experimentally evaluates the performance of both \fips{} and \hips{} by addressing the following questions.
\begin{enumerate}
\item How reliable is the frequency (resp. hyper-volume times frequency) of interval patterns sampled by \fips{} (resp. \hips{})?
\item How does the long-tail phenomenon affect \fips{} and \hips{}?
\item How diverse are the interval patterns sampled by \fips{} and \hips{}?
\item How does \fips{} and \hips{} perform in term of CPU time?
\item What is the relevance of the patterns sampled by \fips{} and \hips{} according to the plausibility criterion?
\end{enumerate}
\textbf{Selected databases.} Our experimental evaluation is conducted on a set of 13 numerical databases. {\it Glass}, {\it Iris}, {\it balance-Scale}, {\it Diabetes}, {\it sonar} and {\it heart} databases are from {the} UCI Machine Learning Repository\footnote{https://archive.ics.uci.edu/}, {while the 7 remaining are from the experimental protocol of our earlier work} \cite{DBLP:conf/cpaior/BekkouchaOBC24}. The number of numerical attributes, objects and distinct values for each database is presented in Table~\ref{table:benchmark}. The results of frequency, plausibility and CPU time experiments are presented for a reduced number of databases. Further results are available in \ref{appendix:expe} and at: { \footnotesize \url{https://github.com/djawed-bkh/Interval-Patterns-Sampling}}
\begin{table}[h]
\centering
\scalebox{1}{
\begin{tabular}{|c||r|r|r|}\hline
Datasets & $\# \M$ & $\# \G$ & $\#$ Distinct values \\\hline
NT & 3 & 130 & 67 \\\hline
AP & 5 & 135 & 674 \\\hline
BK & 5 & 96 & 313 \\\hline
Cancer & 9 & 116 & 900 \\\hline
CH & 8 & 209 & 396 \\\hline
Yacht & 7 & 308 & 322 \\\hline
Iris & 4 & 150 & 123 \\\hline
LW & 10 & 189 & 253 \\\hline
Glass & 9 & 214 & 939 \\\hline
balance-scale & 4 & 625 & 20 \\\hline
diabetes & 8 & 768 & 1254 \\\hline
heart & 13 & 270 & 384 \\\hline
sonar & 60 & 208 & 11 256 \\\hline
\end{tabular}}
\caption{Datasets characteristics}
\label{table:benchmark}
\end{table}
\textbf{Compared approaches.} To the best of our knowledge, there is no approach for sampling interval patterns from numerical data, so we compare \fips{} and \hips{} with a first method that performs a uniform sampling of interval patterns. This first method uniformly draws an attribute $m \in \M$ and then uniformly draws the interval bounds from the values of the attribute $m$. This procedure is repeated until intervals are constructed for all the database attributes. However, as shown in Table~\ref{table:covVide}, this method samples many patterns with empty coverage, making them uninformative for an analyst.
To address this issue, we add a coverage control as described {in} Algorithm~\ref{algo:random} and we call \alea{} the resulting method. While there are available attributes in $\M^*$ (line \ref{uniform:while}), \alea{} uniformly selects and removes an attribute $\im$ from $\M^*$ (lines~\ref{uniform:selectM}-\ref{uniform:dropM}).
{Given the current coverage of the partially constructed interval pattern \( \IP \) (i.e. the subset of objects covered by the intervals already sampled in $\IP$), a candidate set $\N_{m}^{*}$ is defined (line~\ref{uniform:coverNm}).
The interval bounds $[a_m, b_m]$ for attribute $\im$ are then sampled by requiring that at least one bound appears in an object within the coverage of \( \IP \), thereby preventing empty coverage (line~\ref{uniform:bounds}).}
The new sampled interval is added to the partially constructed interval pattern \( \IP \) (line~\ref{uniform:pip}).
Finally, the sampled interval pattern is returned (line \ref{uniform:rip}).
The source code of these methods and experimental results are available at { \footnotesize \url{https://github.com/djawed-bkh/Interval-Patterns-Sampling}}
%Note that the coverage of the partially constructed interval pattern is restricted to the subset of attributes it includes within the database.
\begin{algorithm}[h]
{
\textbf{Input:} $\N$: Numerical database; \\
\textbf{Output:} An interval pattern $\IP$ sampled uniformly with a non empty coverage \;
%\texttt{\\}
$\IP \longleftarrow \langle \rangle$ \tcp*{ Interval pattern under construction}
$\M^* \longleftarrow \{m_1,..., m_{|\M|} \}$\tcp*{Remaining attributes}
\While{$\M^* \neq \emptyset$}{ \label{uniform:while}
$m \longleftarrow$ Draw uniformly an attribute from $\M^*$ \label{uniform:selectM}\;
$\M^* \longleftarrow \M^* \backslash {\{m\}} $ \label{uniform:dropM}\;
$\N_{m}^{*} \longleftarrow \{v_{g,m} \quad |\quad g \in cover(\IP) \}$
\tcp*{ Values from $\N_m$ appearing in objects of the current coverage of $\IP$} \label{uniform:coverNm}
%\textbf{Step 2: }
% Draw uniformly $a_m, b_m$
$[a_m, b_m] \longleftarrow$ Draw uniformly values $a_m, b_m$ such that: $$a_m \leq b_m \text{ and } \left\{\begin{array}{lll}
a_m \in \N_m \land b_m \in \N_{m}^{*} &\\
%\quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \vee
or &\\
a_m \in \N_{m}^{*} \land b_m \in \N_{m} &
\end{array}\right.$$
\tcc*{
selecting from $\N_m^*$ ensures the coverage of at least one object} \label{uniform:bounds}
$\IP \longleftarrow \IP ++ [a_m, b_m]$ \label{uniform:pip}\;
}
\Return $\IP$ \label{uniform:rip}\;
}
\caption{\alea{}: sampling of interval patterns \label{algo:random}}
\end{algorithm}
\begin{table}[h]
\centering
\scalebox{1}{
\begin{tabular}{|c||c|}\hline
Databases & Empty coverage (\%) \\\hline
NT & 8 \\\hline
AP & 76 \\\hline
BK & 66 \\\hline
Cancer & 98 \\\hline
CH & 96 \\\hline
Yacht & 81 \\\hline
Iris & 56 \\\hline
LW & 86 \\\hline
Glass & 99 \\\hline
balance-scale & 0 \\\hline
diabetes & 92 \\\hline
heart & 96 \\\hline
sonar & 100 \\\hline
\end{tabular}}
\caption{Percentage of patterns having an empty coverage when sampling interval patterns according to a uniform distribution without coverage control (10 000 sampled patterns).}
\label{table:covVide}
\end{table}
\subsection{Distribution of Sampled Patterns and Long Tail Phenomenon}
\label{sec:freqandvolume}
Sections~\ref{sec:xp_freq} and \ref{sec:xp_hypervol} address Question 1. We evaluate the patterns sampled by \fips{} and \hips{} according to their respective interestingness measures. These patterns are compared to patterns sampled uniformly using \alea{}. We address Question 2 in Section~\ref{sec:longtaileval} by evaluating the robustness of each approach with regard to the long tail phenomenon.
\subsubsection{Evaluation of Patterns {with Respect to} the Frequency}
\label{sec:xp_freq}
Figure~\ref{exp:frequence} presents the frequencies of 500 patterns sampled by \fips{}, \hips{} and \alea{}. The patterns are sorted in descending order of frequency, and the results are averaged over 10 iterations.
We observe in all databases that \fips{} samples patterns with higher frequencies than those obtained with \alea{}. This can be attributed to Step~1 of \fips{}, which is biased towards selecting objects covered by a large number of patterns, leading to higher frequency. {\hips{} amplifies this frequency bias by jointly considering hyper-volume with frequency into the sampling procedure}.
%In \hips{}, we jointly consider hyper-volume and frequency to ensure that wider intervals also correspond to more frequent patterns. This explains why the patterns produced by \hips{} have higher frequencies than those resulting from the other approaches.
%A sampling procedure based only on hyper-volume tends to bias toward patterns with high hyper-volume. Since an interval pattern with a high hyper-volume is likely to cover more database observations, it is more likely to have a higher frequency. However, a limitation of using hyper-volume alone is that patterns with the same volume would have the same probability of being sampled, regardless of their frequency. However, when considering the product of hyper-volume and frequency, sampling is biased toward patterns that are both wide and frequent. This combination favours patterns that not only cover large regions of the search space, but also include a substantial amount of data observations. This explains the higher frequencies observed with \hips{} than those obtained with \fips{}.
\begin{figure}[h]
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/Frequence/Frequency_evolution_with_hips_7_balance-scale.png}
\end{minipage}
\hspace{0.05\linewidth}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/Frequence/Frequency_evolution_with_hips_8_cancer.png}
\end{minipage}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/Frequence/Frequency_evolution_with_hips_9_diabetes.png}
\end{minipage}
\hspace{0.05\linewidth}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/Frequence/Frequency_evolution_with_hips_6_NT.png}
\end{minipage}
\caption{ Frequency evaluation for 500 patterns sampled with \fips{} and the \alea{} methods }
\label{exp:frequence}
\end{figure}
\subsubsection{Evaluation of Patterns {with Respect to} Hyper-volume $\times$ Frequency}
\label{sec:xp_hypervol}
Figure \ref{exp:volume} shows the values of the product of hyper-volume and frequency of 500 patterns sampled by \fips{}, \hips{} and \alea{}, with patterns sorted in descending order according to this measure. The results are averaged over 10 iterations.
For all databases, \fips{} samples patterns with higher values
compared to those obtained with \alea{}. This is explained by the fact that \fips{} favors frequent patterns. {Since frequent interval patterns cover more database observations}, they are also more likely to have wider intervals, especially when the observations are scattered, resulting in higher volumes and consequently higher values of frequency $\times$ hyper-volume.
For all databases, \hips{} samples patterns with higher values compared to those obtained with \fips{}.
This difference is due to the sampling procedure of \hips{}, which in {Step 1 biases} the object selection towards those covered by patterns with both high frequencies and hyper-volumes and, in Step 2, biases the selection of interval bounds towards wide intervals. Resulting patterns have high frequencies with large hyper-volumes.
\begin{figure}[h]
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/VolumeTimesFrequency/Volume_evolution_with_hips_7_balance-scale.png}
\end{minipage}
\hspace{0.05\linewidth}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/VolumeTimesFrequency/Volume_evolution_with_hips_8_cancer.png}
\end{minipage}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/VolumeTimesFrequency/Volume_evolution_with_hips_9_diabetes.png}
\end{minipage}
\hspace{0.05\linewidth}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/VolumeTimesFrequency/Volume_evolution_with_hips_6_NT.png}
\end{minipage}
\caption{ Volume evaluation for 500 patterns sampled with \hips{} and the \alea{} methods }
\label{exp:volume}
\end{figure}
%\subsection{Pattern Frequencies and Long Tail Phenomenon}
%\label{sec:frequency}
\subsubsection{Long Tail Phenomenon}
\label{sec:longtaileval}
Question~2 related to the long tail phenomenon is addressed in this section. The \textit{long tail phenomenon} is a statistical effect introduced { by Bryson~\cite{Bryson01021974} and popularized in the context of business by Anderson~\cite{anderson2006longtail}. It affects sampling techniques by leading to the sampling of uninteresting patterns.}
The long tail is characterized by an unbalanced distribution of sampled patterns with respect to the considered measure. A small number of patterns with high measure values form the {``head''}, while a large number of patterns with low measure values form the {``tail''}. In the following, we evaluate the robustness of \fips{} and \hips{} to the long tail phenomenon, each one according to its respective interestingness measure, namely frequency for \fips{} and the product of frequency and hyper-volume for \hips{}.
\vspace*{0.2cm}
\textbf{Robustness of \fips{} to the long tail phenomenon. }
Figure \ref{exp:frequence} shows that \fips{} is significantly less affected by the long-tail phenomenon than \alea{}. For the {\it diabetes} and {\it cancer} databases, 65\% and 68.2\% of patterns generated by \fips{} occur in the tail (frequency below 1\%), in comparison to approximately 99\% with \alea{}. For the {\it balance-scale} and {\it NT} databases, \fips{} produces a negligible number of low-frequency patterns, making the tail almost undetectable. In contrast, \alea{} produces 9.4\% and 10\% of patterns with frequencies below 1\% for the balance-scale and NT databases, respectively.
\vspace*{0.2cm}
\textbf{Robustness of \hips{} to the long tail phenomenon} Similarly to \fips{}, Figure~\ref{exp:volume} shows that \hips{} is significantly less affected by the long-tail phenomenon than \alea{}, with respect to the considered interestingness measure. In the \textit{NT} database, the tail is almost undetectable. In the \textit{balance-scale}, \textit{cancer}, and \textit{diabetes} datasets, we observe that \hips{} consistently delays the appearance of the tail. For instance, in \textit{balance-scale}, only 50\% of the sampled patterns fall within the tail, compared to 86\% for \alea{}. In \textit{cancer}, the difference is even more pronounced: 85\% for \hips{} versus 99\% for \alea{}. Similarly, in \textit{diabetes}, 60\% of the \hips{} patterns appear in the tail, compared to 99\% for \alea{}.
%This result is mainly due to the sampling procedure of \hips{}, which, in its first step, tends to draw objects that are covered by patterns with both high frequency and large volume. In its second step, \hips{} further biases the sampling toward patterns with wide intervals. These two steps combined contribute to delaying, or even suppressing, the emergence of the tail.
%For the {\it diabetes} and {\it cancer} databases, 65\% and 68.2\% of patterns generated by \fips{} occur in the tail (frequency below 1\%), in comparison to approximately 99\% with \alea{}. For the {\it balance-scale} and {\it NT} databases, \fips{} produces a negligible number of low-frequency patterns, making the tail almost undetectable. In contrast, the \alea{} method produces 9.4\% and 10\% of patterns with frequencies below 1\% for the balance-scale and NT databases, respectively.
%For the frequency measure, Figure \ref{exp:frequence} shows that \hips{} is less affected by the long-tail phenomenon than \fips{} and \alea{}. In the majority of the databases, the tail is almost undetectable. This can be attributed to the fact that \hips{} draws patterns with high volumes, which lead to higher frequencies. For the \fips{} method, we observe that it is significantly less affected by the long-tail phenomenon than \alea{} in the \textit{NT} and \textit{balance-scale} databases, where the tail is difficult to detect. In contrast, for the \textit{cancer} and \textit{diabetes} databases, the results are closer between the two methods, although \fips{} manages to slightly delay the appearance of the tail. This can be explained by the number of attributes in these databases (see Table \ref{table:benchmark}). The larger the database in terms of the number of attributes, the smaller the overlap between intervals, which results in a lower frequency.
%For the volume measure, we observe similar trends in Figure \ref{exp:volume}, with \hips{} being the least vulnerable to the long-tail phenomenon. \fips{} is also less affected in the \textit{balance scale} and \textit{NT} databases, while the results for the \textit{cancer} and \textit{diabetes} databases are closer to \alea{}.
%\todo[inline]{DB: est ce que je garde ce graphique, ou je mets les memes graphiques que j'ai mis dans l'évaluation du volume et de la fréquence (affichés plus en haut), mais cette fois ci avec juste $HFIPS$ et $FIPS$ }
%\textbf{\fips{} and \hips{} comparison.} Since both \fips{} and \hips{} are sampling methods that rely, at least partially, on frequency as an interestigness measure (\fips{} uses frequency measure, while \hips{} uses frequency multiplied by volume), it is interesting to compare the two approaches to assess the impact of incorporating volume into the sampling measure. Thus, we evaluate both the frequency of patterns sampled by \fips{} and \hips{}, as well as the hyper-volumes obtained with each approach. In figure \ref{fig:hipsAndFipsComparison} we compare the average frequencies and hyper-volumes of 1000 patterns sampled by \fips{} and \hips{}. For frequency, \hips{} samples patterns with a higher average frequency than \fips{}. This can be attributed to the sampling measure used in \hips{}. Sampling patterns based on the product of frequency and hyper-volume biases towards patterns with both high frequencies and large hyper-volumes. This is intuitive since larger patterns are more likely to include a greater number of observations from the dataset.
%This behaviour is further explained by the algorithm \ref{algo:hv_sampling}. In step 2, the sampling procedure is biased towards selecting bounds that form large intervals. More specifically, there is a higher probability of selecting the smallest distinct values of the attribute as lower bounds and the largest as upper bounds (lines 5 and 6 of algorithm \ref{algo:hv_sampling}). In contrast, \fips{} uniformly selects patterns from the set of patterns covering at least the sampled object in step 1 of algorithm \ref{algo:FIPS}. A similar trend is observed for hyper-volume: \hips{} produces patterns with larger hyper-volumes than \fips{}, for the same reasons as for frequency.
%\begin{figure}[h]
% \centering
% \includegraphics[width=\linewidth]{Images/mean_freq_and_volume_comparison_with_stdev.png}
% \caption{Comparison of mean frequencies and hyper-volumes of 1000 interval patterns sampled by \fips{} and \hips{}}
% \label{fig:hipsAndFipsComparison}
%\end{figure}
\subsection{Diversity Evaluation}
Sections~\ref{sec:eqclassdiversity} and \ref{sec:jaccardDIversity} address Question 3 on the diversity of sampled patterns. Sampling diverse patterns is essential in an interactive mining process, as it allows the analyst to explore multiple areas of the solution space. In this section, we use various metrics to assess the diversity of patterns sampled by \fips{}, \hips{} and \alea{}. Section~\ref{sec:eqclassdiversity} deals with the diversity of equivalence classes among the sampled patterns, while Section~\ref{sec:jaccardDIversity} focuses on their overlap in terms of covered objects.
\subsubsection{Equivalence Classes Diversity}
\label{sec:eqclassdiversity}
To evaluate the diversity of \fips{}, \hips{} and \alea{} in terms of equivalence classes, we use the measure presented in \cite{DBLP:conf/sdm/GiacomettiS18}, which is defined as follows: $$ diversity(K, \N) = \frac{|\{cover(\IP_1, \N), \dots, cover(\IP_{|K|}, \N)\}|}{|K|}$$ where $|K|$ is the number of sampled patterns. Figure~\ref{exp:DiversitySDM} shows that patterns sampled with \hips{} and \fips{} have a greater diversity in equivalence classes than those sampled with \alea{}, except in the case of the {\it balance-scale} database. This can be attributed to the higher frequency of patterns produced by \hips{} and \fips{} (see Section~\ref{sec:freqandvolume}), which results in wider coverage and, consequently, greater variation in the covered objects across samples, thereby increasing diversity in terms of equivalence classes. The exception observed for the {\it balance-scale} dataset can be explained by its specific characteristics, in particular the limited number of distinct values and numerical attributes (see Table~\ref{table:benchmark}). Since \fips{} and \hips{} are biased towards high frequency and high hyper-volume patterns, they tend to sample the same objects repeatedly, resulting in similar intervals and reduced diversity.
%Due to the volume and frequency biased sampling of \hips{} and \fips{}, the same objects and thus similar intervals are sampled.
%Regarding the higher diversity of \hips{} compared to \fips{}, this can be explained by the fact that \hips{} samples patterns with higher frequencies than \fips{}, which lead to a higher variation in the objects covered by the patterns sampled by \hips{} compared to \fips{}.
The higher diversity of \hips{} in comparison to \fips{} can be explained by the higher frequencies of patterns sampled by \hips{}, resulting in greater variation in the coverages. %covered objects.
%Sampling diverse patterns is essential in an interactive mining process, as it allows the analyst to explore multiple areas of the solution space. To evaluate the diversity of \fips{} and \alea{}, we use the measure presented in \cite{DBLP:conf/sdm/GiacomettiS18}, which is tailored for numerical data and defined as follows: $ diversity(K, \N) = |\{cover(\IP_1, \N), \dots, cover(\IP_{|K|}, \N)\}|$ $/|K|$ where $|K|$ is the number of sampled patterns. Figure~\ref{exp:DiversitySDM} shows that patterns sampled with \fips{} have greater diversity than those sampled with the \alea{} method, except in the case of the {\it balance-scale} database. This can be explained by the frequency of patterns obtained through \fips{} which tends to be higher than those obtained with the \alea{} method (see Section \ref{sec:frequency}). This results in higher coverage and, consequently, a greater probability of variation in coverage across samples. In the case of the {\it balance-scale} database, the occurring exception can be attributed to the specific characteristics of this database, which has few distinct values and numerical attributes (see Table \ref{table:benchmark}). Due to the frequency-biased sampling of \fips{}, the same objects and thus similar intervals are sampled.
\begin{figure}[h]
\centering
\includegraphics[width=1\linewidth]{Images/Diversity/Diversity_soulet_bar_plotEN.png}
\caption{Diversity evaluation of \fips{}, \hips ~and \alea{} methods}
\label{exp:DiversitySDM}
\end{figure}
\subsubsection{Coverage Overlapping}
\label{sec:jaccardDIversity}
Since the metric presented in Section \ref{sec:eqclassdiversity} does not consider the overlap of coverages between the sampled patterns, we complement the diversity evaluation %of the \hips{}, \fips{} and \alea{} approaches
with a Jaccard index-based measure. We evaluate the coverage diversity of 500 patterns obtained with \fips{}, \hips{} and \alea{}, using the Jaccard index. As this metric compares pairs of patterns, it cannot be applied directly to a large set. Therefore, we use cumulative distribution functions (CDF) to visually represent the distribution of Jaccard indices computed for all pairs formed by our 500 patterns. Consider a set of \( k=500 \) interval patterns \( \boldsymbol{\IP}=\{\IP_1, \dots, \IP_{500}\} \) and a numerical database \( \N \).
\[
CDF(\N, \boldsymbol{\IP}, \theta) = \frac{ 2 \times \left| \{(i,j) : Jac(\IP_i, \IP_j) \leq \theta ,~~ 1 \leq i < j \leq k,~~ \IP_i, \IP_j \in \boldsymbol{\IP} \} \right|}{k(k-1)}
\]
Figure~\ref{exp:jaccardDiversity} shows that, for all databases, the cumulative distribution function curve of \hips{} has a lower area compared to \alea{}. This indicates that \hips{} generates patterns with higher similarity in terms of coverage than \alea{}. For small Jaccard indices, \hips{} returns a small number of patterns with low coverage similarity. However, as the similarity threshold increases, the number of patterns produced by \hips{} also grows, meaning that this method generates patterns that share the same objects in their coverage. In contrast, \alea{} shows greater coverage diversity: a large number of patterns have low Jaccard indices, and the curve generally remains linear as the similarity threshold increases. This means that there is no significant addition of highly similar patterns. This observation can be explained by the fact that \alea{} tends to sample patterns covering a small number of objects, reducing the probability that these patterns share objects with others from the same approach. In contrast, \hips{}, which samples patterns covering a large number of objects, increases the probability of sampling other patterns that share a subset of objects.
For the \fips{} method, we observe that for the \textit{NT} and \textit{balance-scale} databases, the cumulative distribution function curve has a lower area compared to the random sampling approach. This indicates that \fips{} generates patterns with higher similarity in terms of coverage than \alea{}. This can be explained by two factors: first, \fips{} samples patterns with greater frequencies than \alea{} (see Figure \ref{sec:freqandvolume}), meaning that patterns obtained with \fips{} are more likely to share the same objects in their coverages, leading to lower diversity. The second factor concerns the database characteristics (see Table \ref{table:benchmark}). We observe that \textit{NT} and \textit{balance-scale} have a small number of distinct values, which leads \fips{} to often sample the same patterns.
Finally, for the \textit{diabetes} and \textit{cancer} databases, we observe that the diversity of \fips{} is almost as high as that of \alea{}. This can be explained by the large number of distinct values in these databases, which leads \fips{} to sample more diversified patterns.
\begin{figure}[h]
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/Diversity/Jaccard/Jaccard_CDF_7_balance-scale.png}
\end{minipage}
\hspace{0.05\linewidth}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/Diversity/Jaccard/Jaccard_CDF_8_cancer.png}
\end{minipage}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/Diversity/Jaccard/Jaccard_CDF_9_diabetes.png}
\end{minipage}
\hspace{0.05\linewidth}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/Diversity/Jaccard/Jaccard_CDF_6_NT.png}
\end{minipage}
\caption{ Coverage diversity evaluation for 500 patterns sampled with \fips{}, \hips ~and the \alea{} method }
\label{exp:jaccardDiversity}
\end{figure}
%This allows a better understanding of how similarities between patterns are distributed according to their Jaccard index.
\subsection{CPU Time Evaluation}
This section deals with running time (Question 4). Figure~\ref{exp:tempsCPU} presents the CPU time\footnote{Configuration: Intel Core i7, 11th generation, 3 GHz, 8 cores, 30 GB RAM} required to sample each of 500 interval patterns using \fips{}, \hips{}, and \alea{}. The results are averaged over 10 iterations.
We observe that for most of the databases, \fips{} is faster than both \alea{} and \hips{}. This can be explained by the fact that \alea{} requires, for each sampled interval in the pattern, a search for distinct values from the current coverage. This is achieved by recalculating the current coverage of the pattern being constructed for each newly sampled interval ($|\M|$ times in total). This procedure also explains the oscillations observed for this method. If the gap between the bounds of the first sampled intervals is large, the coverage computation time will be high. In contrast, if the gap is small, the computation time will be lower. With the exception of \textit{balance-scale}, this time is longer for \hips{} than for \alea{} and \fips{}. This is due to the second step of its algorithm (see Algorithm~\ref{algo:hv_sampling}), where the weight computation of the values that are candidates for being bound takes more CPU time. The exception on \textit{balance-scale} can be explained by the small number of distinct values present in this database (see Table~\ref{table:benchmark}), which results in a smaller number of candidate values for being bounds, thus requiring less computation time for the second step of \hips{} (see Algorithm~\ref{algo:hv_sampling}). Note that in {all cases}, run-times remain very short as the time scale is measured in milliseconds.
\begin{figure}[h]
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/TempsCPU/CPU_evolution_comparison_7_balance-scale.png}
\end{minipage}
\hspace{0.05\linewidth}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/TempsCPU/CPU_evolution_comparison_8_cancer.png}
\end{minipage}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/TempsCPU/CPU_evolution_comparison_9_diabetes.png}
\end{minipage}
\hspace{0.05\linewidth}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/TempsCPU/CPU_evolution_comparison_6_NT.png}
\end{minipage}
\caption{
CPU time evolution for a set of 500 patterns sampled by the \fips{}, \hips{} and \alea{} methods }
\label{exp:tempsCPU}
\end{figure}
\subsection{Plausibility Evaluation}
This section addresses Question 5 on the relevance of the patterns sampled by \fips{}, \hips{} and \alea{} according to the plausibility criterion. We use the protocol introduced by \cite{DBLP:conf/kdd/GionisMMT06} and adapted to numerical data in \cite{DBLP:conf/sdm/GiacomettiS18}.
The idea is to evaluate the relevance of a pattern by comparing its frequency in the original database with its frequency in a randomized database, denoted $\N_{rand}$, in which correlations are broken. {This randomization is performed by repeatedly selecting at random, from the original database, two values of the same attribute from two different objects and swapping them. This process is applied independently for each attribute, and a random number of swaps is performed. A pattern’s relevance increases with the difference between its frequency in the original database and in randomized variants.} Formally, plausibility is defined as : $$Plausibility(K, \N)= \frac{\sum_{\IP_i \in {K}}\sum_{j=1}^{R} (freq(\IP_i,~ \N) - freq(\IP_i,~ \N_{rand}^{j}))} {\sum_{i=1}^{|K|} (R \times freq(\IP_i, \N))}$$ where $R$ is the number of randomised databases.
%The smaller a pattern's frequency in the randomized database, the more relevant it is considered.
%
We defined frequency intervals (see Figure~\ref{fig:plausibility}) and for each interval and method we sampled 10,000 patterns, rejecting those whose frequencies {did not fall within the specified range}.
The restriction imposed by the frequency intervals aims to simulate the interest of an analyst who rejects both highly frequent patterns (often too general) and very rare patterns (often not representative).
Considering multiple intervals allow us to simulate several thresholds values considered useful. Sampling was conducted within a time limit of 5 minutes.
\begin{figure}[h]
\centering
\includegraphics[width=1\linewidth]{Images/Plausibility/Plausibility_et_Nombre_de_tirages_fips_vs_random.png}
\caption{Plausibility evaluation for the \fips{}, \hips{} and \alea{} methods (10 000 sampled patterns) on multiple frequency thresholds}
\label{fig:plausibility}
\end{figure}
\vspace*{0.2cm}
\textbf{Plausibility. } Plausibility tends to decrease as frequency increases, as the more frequent a pattern is, the more likely it is to appear in a randomized database. Since the \fips{} and \hips{} sampling are biased towards frequency and the product of hyper-volume and frequency, respectively, patterns sampled by these methods are expected to have lower plausibility on average than those sampled by \alea{}. Part A of Figure~\ref{fig:plausibility} confirms this observation. As \alea{} produces patterns with the lowest frequencies, their plausibility is the highest. Conversely, \hips{}, which tends to sample the most frequent patterns, leads to patterns with the lowest plausibility. The results also show that as the frequency threshold increases, the plausibility of the patterns sampled by the different methods converges. Note that some methods fail to sample the required number of patterns within the allocated time. For instance, \alea{} fails to sample the required number of patterns with frequencies greater than 30\% for the \textit{AP} and \textit{BK} databases. On the other hand, \hips{} fails to sample the required number of patterns in \textit{AP} with a frequency lower than 45\% within the allocated time.
\vspace*{0.2cm}
\textbf{Number of draws evaluation.} Parts $B_1$, $B_2$, $B_3$ and $B_4$ in Figure~\ref{fig:plausibility} show that across all databases, the number of draws required by \alea{} to reach the target number of patterns is significantly higher than that of \fips{} and \hips{}. For frequency thresholds between 10\% and 35\%, \alea{} requires 0.87 to 10 times more draws than \hips{} on {\it iris}, 1.04 to 7 times more draws on {\it balance-scale}, and 2 to 11 times more draws than \fips{} on both {\it iris} and {\it balance-scale}. This gap increases for higher frequency thresholds (35\% to 45\%), reaching up to 27 times more draws than \fips{} and 35 times more draws than \hips{} on the {\it iris} database. For \fips{}, the difference is even more pronounced on the {\it AP} and {\it BK} databases, where \alea{} requires 15 to 80 times more draws in low frequency thresholds. At higher thresholds, \alea{} fails to sample the required number of patterns in the allocated time. Similarly, for \hips{}, the gap is particularly pronounced on the {\it BK} and {\it iris} databases, where \alea{} requires up to 100 times more draws, and reaches 134 times more on {\it iris}.
\section{Conclusion and Perspectives }
In this paper, we presented two methods for sampling patterns from numerical data. These methods rely on an interval-based representation of the numerical data that preserves all the original information. The first method, \fips{}, samples patterns proportionally to the frequency interestingness measure. % We extended \fips{} with \hips{}, %which incorporates the hyper-volume into the sampling process, allowing \hips{} to sample
\hips{} incorporates the hyper-volume measure into the sampling process in order to sample interval patterns proportionally to the product of frequency and hyper-volume. We theoretically proved that both \fips{} and \hips{} sample patterns proportionally to their respective interestingness measures. We experimentally evaluated the quality of the sampled patterns using several criteria, including frequency, hyper-volume, diversity, efficiency, and plausibility. We also evaluate the robustness of our approaches to the long-tail phenomenon.
Increasingly, analysts want to interact with mining systems by specifying, through queries, the types of patterns they want to explore. In this context, a future work is to design a framework for sampling interval patterns that satisfy user preferences and follows the desired distribution.
This task is particularly challenging since, unlike itemsets, interval patterns are defined across all attributes of the database and include hidden dependencies between intervals. Another direction is to tackle other interestingness measures that are relevant to numerical data, such as density. Finally, exploring the use of sampling in machine learning contexts could represent a promising direction. For example, sampling approaches could help identify meaningful hyper-parameters for supervised learning tasks. By guiding the search toward informative regions of the hyper-parameter space, our approaches may help to reduce computational costs and improve the explainability of the resulting models, which would be a step toward more interpretable machine learning models.
\vspace*{0.2cm}
\noindent
\textbf{Acknowledgement.} This work was supported by the French National Research Agency (ANR) and Region Normandie under grant HAISCoDe, and partly funded by ANR under project ANR-24-CE23-0950.
%\appendix
\section{Proof of the $IPH$ function}
\label{appendix:tilProof}
% Preuve proposée par Lamine
% \begin{proof}
% Assume that the distinct values of attribute $m$ are given in sorted order by
% \[
% \mathcal{D}_m = \{d_1, d_2, \dots, d_k\} \quad \text{with} \quad d_1 < d_2 < \cdots < d_k.
% \]
% Let $p$ be defined as
% \[
% p = \max\{i \mid d_i \leq v\}.
% \]
% Then the candidate lower and upper bound sets can be written as
% \[
% I_{g,m} = \{d_1, d_2, \dots, d_p\} \quad \text{and} \quad J_{g,m} = \{d_p, d_{p+1}, \dots, d_k\}.
% \]
%
% Intuitively, the Total Interval Area sums the lengths of all intervals $[a, b]$ that cover the value $v$, where $a \in I_{g,m}$ and $b \in J_{g,m}$. Formally, this is written as:
% \[
% TIA_m(v) = \sum_{a \in I_{g,m}} \sum_{b \in J_{g,m}} (b - a).
% \]
% Substituting the ordered sets, we have:
% \[
% TIA_m(v) = \sum_{i=1}^{p} \sum_{j=p}^{k} (d_j - d_i).
% \]
% This double summation can be split into two parts:
% \[
% TIA_m(v) = \underbrace{\sum_{i=1}^{p} \sum_{j=p}^{k} d_j}_{\text{Term } A} - \underbrace{\sum_{i=1}^{p} \sum_{j=p}^{k} d_i}_{\text{Term } B}.
% \]
% Consider Term $A$: The inner sum $\sum_{j=p}^{k} d_j$ is independent of the index $i$, so
% \[
% \sum_{i=1}^{p} \sum_{j=p}^{k} d_j = p \cdot \left(\sum_{j=p}^{k} d_j\right).
% \]
% For Term $B$: The inner sum $\sum_{j=p}^{k} d_i$ is independent of the index $j$, and since there are $|J_{g,m}| = k - p + 1$ terms for each $i$, we have
% \[
% \sum_{i=1}^{p} \sum_{j=p}^{k} d_i = \sum_{i=1}^{p} \left[(k-p+1) \, d_i\right] = (k-p+1) \cdot \left(\sum_{i=1}^{p} d_i\right).
% \]
% Thus, combining the two terms, we obtain:
% \[
% TIL_m(v) = p \cdot \sum_{j=p}^{k} d_j - (k-p+1) \cdot \sum_{i=1}^{p} d_i.
% \]
% Recognizing that $p = |I_{g,m}|$ and $k-p+1 = |J_{g,m}|$, we can rewrite this as:
% \[
% TIL_m(v) = |I_{g,m}| \cdot \sum_{x \in J_{g,m}} x - |J_{g,m}| \cdot \sum_{x \in I_{g,m}} x.
% \]
% This completes the proof.
% \end{proof}
% DB: j'ai repris la preuve de Lamine en considérant le TIL comme étant un terme de la fonction IPH, donc la preuve concerne IPH maintenant
\begin{proof}
We aim to prove that the $IPH$ function correctly computes the sum of the hyper-volumes of all interval patterns that cover a given object $g \in \G$. This requires demonstrating that, for any attribute $m \in \M$, the inner term of Equation~\ref{eq:IPH} corresponds to the sum of the lengths of all intervals that contain the value $\val{}{}$ of $g$ on attribute $m$.
Assume that the distinct values of the $m$ attribute are given in a sorted order by:
\[
\mathcal{D}_m = \{d_1, d_2, \dots, d_k\}, \quad \text{with} \quad d_1 < d_2 < \cdots < d_k.
\]
Let $p$ be the index of the highest value in $\mathcal{D}_m$ being less than or equal to $\val{}{}$:
\[
p = \max\{i \mid d_i \leq \val{}{}\}.
\]
Then the candidate sets of lower and upper bounds for an attribute $m$ can be defined as follows:
\[
\I{\val{}{}} = \{d_1, \dots, d_p\}, \qquad \J{\val{}{}} = \{d_p, \dots, d_k\}.
\]
The total interval length $TIL(\val{}{})$ corresponds to the sum of the lengths of all intervals $[a, b]$ such that $a \in I_{g,m}$, $b \in J_{g,m}$, and $a \leq \val{}{} \leq b$. Formally, we write:
\[
\TIL{m}{g} = \sum_{a \in I_{g,m}} \sum_{b \in J_{g,m}} (b - a).
\]
By substituting with the ordered sets, we have:
\[
{\TIL{m}{g}} = \sum_{i=1}^{p} \sum_{j=p}^{k} (d_j - d_i).
\]
By expanding this double summation, we obtain:
\[
\TIL{m}{g} = \sum_{i=1}^{p} \sum_{j=p}^{k} (d_j - d_i) = \sum_{i=1}^{p} \sum_{j=p}^{k} d_j - \sum_{i=1}^{p} \sum_{j=p}^{k} d_i.
\]
Since the inner sum is independent of $i$ and repeated $p$ times, the first term can be factored as:
\[
\sum_{i=1}^{p} \sum_{j=p}^{k} d_j = p \cdot \sum_{j=p}^{k} d_j,
\]
Similarly, since the inner sum is independent of $j$ and repeated $k - p + 1$ times, the second term becomes:
\[
\sum_{i=1}^{p} \sum_{j=p}^{k} d_i = (k - p + 1) \cdot \sum_{i=1}^{p} d_i,
\]
Recognizing that $p = |\I{\val{}{}}|$ and $k-p+1 = |\J{\val{}{}}|$, we can rewrite this as:
\[
\TIL{m}{g} = |\I{\val{}{}}| \cdot \sum_{x \in \J{\val{}{}}} x - |\J{\val{}{}}| \cdot \sum_{x \in \I{\val{}{}}} x.
\]
which matches the expression given in Equation~\ref{eq:IPH} for an attribute $m$. Since the hyper-volume of an interval pattern is the product of the lengths of its intervals, applying the product over all attributes leads to the total hyper-volume of all interval patterns covering object $g$.
\end{proof}
\section{Experimental results}
\label{appendix:expe}
\subsection{Pattern Frequencies Evaluation}
\begin{figure}[H]
\centering
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/Frequence/Frequency_evolution_with_hips_1_AP.png}
\end{minipage}
\hspace{0.05\linewidth}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/Frequence/Frequency_evolution_with_hips_2_BK.png}
\end{minipage}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/Frequence/Frequency_evolution_with_hips_3_CH.png}
\end{minipage}
\hspace{0.05\linewidth}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/Frequence/Frequency_evolution_with_hips_4_Iris.png}
\end{minipage}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/Frequence/Frequency_evolution_with_hips_5_LW.png}
\end{minipage}
\hspace{0.05\linewidth}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/Frequence/Frequency_evolution_with_hips_10_glass.png}
\end{minipage}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/Frequence/Frequency_evolution_with_hips_11_heart.png}
\end{minipage}
\hspace{0.05\linewidth}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/Frequence/Frequency_evolution_with_hips_12_sonar.png}
\end{minipage}
\end{figure}
\begin{figure}[H]
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/Frequence/Frequency_evolution_with_hips_13_yacht.png}
\end{minipage}
\end{figure}
\subsection{Pattern Volume Times Frequency Evaluation}
\begin{figure}[H]
\centering
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/VolumeTimesFrequency/Volume_evolution_with_hips_1_AP.png}
\end{minipage}
\hspace{0.05\linewidth}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/VolumeTimesFrequency/Volume_evolution_with_hips_2_BK.png}
\end{minipage}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/VolumeTimesFrequency/Volume_evolution_with_hips_3_CH.png}
\end{minipage}
\hspace{0.05\linewidth}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/VolumeTimesFrequency/Volume_evolution_with_hips_4_Iris.png}
\end{minipage}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/VolumeTimesFrequency/Volume_evolution_with_hips_2_BK.png}
\end{minipage}
\hspace{0.05\linewidth}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/VolumeTimesFrequency/Volume_evolution_with_hips_10_glass.png}
\end{minipage}
\end{figure}
\begin{figure}[H]
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/VolumeTimesFrequency/Volume_evolution_with_hips_11_heart.png}
\end{minipage}
\hspace{0.05\linewidth}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/VolumeTimesFrequency/Volume_evolution_with_hips_12_sonar.png}
\end{minipage}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/VolumeTimesFrequency/Volume_evolution_with_hips_13_yacht.png}
\end{minipage}
\end{figure}
\subsection{Jaccard based diversity evaluation}
\begin{figure}[H]
\centering
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/Diversity/Jaccard/Jaccard_CDF_1_AP.png}
\end{minipage}
\hspace{0.05\linewidth}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/Diversity/Jaccard/Jaccard_CDF_2_BK.png}
\end{minipage}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/Diversity/Jaccard/Jaccard_CDF_3_CH.png}
\end{minipage}
\hspace{0.05\linewidth}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/Diversity/Jaccard/Jaccard_CDF_4_Iris.png}
\end{minipage}
\end{figure}
\begin{figure}[H]
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/Diversity/Jaccard/Jaccard_CDF_5_LW.png}
\end{minipage}
\hspace{0.05\linewidth}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/Diversity/Jaccard/Jaccard_CDF_10_glass.png}
\end{minipage}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/Diversity/Jaccard/Jaccard_CDF_11_heart.png}
\end{minipage}
\hspace{0.05\linewidth}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/Diversity/Jaccard/Jaccard_CDF_12_sonar.png}
\end{minipage}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/Diversity/Jaccard/Jaccard_CDF_13_yacht.png}
\end{minipage}
\end{figure}
\subsection{CPU time evaluation}
\begin{figure}[H]
\centering
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/TempsCPU/CPU_evolution_comparison_1_AP.png}
\end{minipage}
\hspace{0.05\linewidth}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/TempsCPU/CPU_evolution_comparison_2_BK.png}
\end{minipage}
\end{figure}
\begin{figure}[H]
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/TempsCPU/CPU_evolution_comparison_3_CH.png}
\end{minipage}
\hspace{0.05\linewidth}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/TempsCPU/CPU_evolution_comparison_4_Iris.png}
\end{minipage}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/TempsCPU/CPU_evolution_comparison_5_LW.png}
\end{minipage}
\hspace{0.05\linewidth}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/TempsCPU/CPU_evolution_comparison_10_glass.png}
\end{minipage}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/TempsCPU/CPU_evolution_comparison_11_heart.png}
\end{minipage}
\hspace{0.05\linewidth}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/TempsCPU/CPU_evolution_comparison_12_sonar.png}
\end{minipage}
\begin{minipage}[b]{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{Images/TempsCPU/CPU_evolution_comparison_13_yacht.png}
\end{minipage}
\end{figure}