Plausible Counterfactual Explanations of Recommendations
Extracted source text · arXiv 2507.07919
%% bare_conf_compsoc.tex
%% V1.4b
%% 2015/08/26
%% by Michael Shell
%% See:
%% http://www.michaelshell.org/
%% for current contact information.
%%
%% This is a skeleton file demonstrating the use of IEEEtran.cls
%% (requires IEEEtran.cls version 1.8b or later) with an IEEE Computer
%% Society conference paper.
%%
%% Support sites:
%% http://www.michaelshell.org/tex/ieeetran/
%% http://www.ctan.org/pkg/ieeetran
%% and
%% http://www.ieee.org/
%%*************************************************************************
%% Legal Notice:
%% This code is offered as-is without any warranty either expressed or
%% implied; without even the implied warranty of MERCHANTABILITY or
%% FITNESS FOR A PARTICULAR PURPOSE!
%% User assumes all risk.
%% In no event shall the IEEE or any contributor to this code be liable for
%% any damages or losses, including, but not limited to, incidental,
%% consequential, or any other damages, resulting from the use or misuse
%% of any information contained here.
%%
%% All comments are the opinions of their respective authors and are not
%% necessarily endorsed by the IEEE.
%%
%% This work is distributed under the LaTeX Project Public License (LPPL)
%% ( http://www.latex-project.org/ ) version 1.3, and may be freely used,
%% distributed and modified. A copy of the LPPL, version 1.3, is included
%% in the base LaTeX documentation of all distributions of LaTeX released
%% 2003/12/01 or later.
%% Retain all contribution notices and credits.
%% ** Modified files should be clearly indicated as such, including **
%% ** renaming them and changing author support contact information. **
%%*************************************************************************
% *** Authors should verify (and, if needed, correct) their LaTeX system ***
% *** with the testflow diagnostic prior to trusting their LaTeX platform ***
% *** with production work. The IEEE's font choices and paper sizes can ***
% *** trigger bugs that do not appear when using other class files. *** ***
% The testflow support page is at:
% http://www.michaelshell.org/tex/testflow/
% \documentclass[conference,compsoc]{IEEEtran}
\documentclass[conference]{IEEEtran}
% Some/most Computer Society conferences require the compsoc mode option,
% but others may want the standard conference format.
%
% If IEEEtran.cls has not been installed into the LaTeX system files,
% manually specify the path to it like:
% \documentclass[conference,compsoc]{../sty/IEEEtran}
% Some very useful LaTeX packages include:
% (uncomment the ones you want to load)
% *** MISC UTILITY PACKAGES ***
%
%\usepackage{ifpdf}
% Heiko Oberdiek's ifpdf.sty is very useful if you need conditional
% compilation based on whether the output is pdf or dvi.
% usage:
% \ifpdf
% % pdf code
% \else
% % dvi code
% \fi
% The latest version of ifpdf.sty can be obtained from:
% http://www.ctan.org/pkg/ifpdf
% Also, note that IEEEtran.cls V1.7 and later provides a builtin
% \ifCLASSINFOpdf conditional that works the same way.
% When switching from latex to pdflatex and vice-versa, the compiler may
% have to be run twice to clear warning/error messages.
% *** CITATION PACKAGES ***
%
\ifCLASSOPTIONcompsoc
% IEEE Computer Society needs nocompress option
% requires cite.sty v4.0 or later (November 2003)
\usepackage[nocompress]{cite}
\else
% normal IEEE
\usepackage{cite}
\fi
% cite.sty was written by Donald Arseneau
% V1.6 and later of IEEEtran pre-defines the format of the cite.sty package
% \cite{} output to follow that of the IEEE. Loading the cite package will
% result in citation numbers being automatically sorted and properly
% "compressed/ranged". e.g., [1], [9], [2], [7], [5], [6] without using
% cite.sty will become [1], [2], [5]--[7], [9] using cite.sty. cite.sty's
% \cite will automatically add leading space, if needed. Use cite.sty's
% noadjust option (cite.sty V3.8 and later) if you want to turn this off
% such as if a citation ever needs to be enclosed in parenthesis.
% cite.sty is already installed on most LaTeX systems. Be sure and use
% version 5.0 (2009-03-20) and later if using hyperref.sty.
% The latest version can be obtained at:
% http://www.ctan.org/pkg/cite
% The documentation is contained in the cite.sty file itself.
%
% Note that some packages require special options to format as the Computer
% Society requires. In particular, Computer Society papers do not use
% compressed citation ranges as is done in typical IEEE papers
% (e.g., [1]-[4]). Instead, they list every citation separately in order
% (e.g., [1], [2], [3], [4]). To get the latter we need to load the cite
% package with the nocompress option which is supported by cite.sty v4.0
% and later.
% *** GRAPHICS RELATED PACKAGES ***
%
\ifCLASSINFOpdf
% \usepackage[pdftex]{graphicx}
% declare the path(s) where your graphic files are
% \graphicspath{{../pdf/}{../jpeg/}}
% and their extensions so you won't have to specify these with
% every instance of \includegraphics
% \DeclareGraphicsExtensions{.pdf,.jpeg,.png}
\else
% or other class option (dvipsone, dvipdf, if not using dvips). graphicx
% will default to the driver specified in the system graphics.cfg if no
% driver is specified.
% \usepackage[dvips]{graphicx}
% declare the path(s) where your graphic files are
% \graphicspath{{../eps/}}
% and their extensions so you won't have to specify these with
% every instance of \includegraphics
% \DeclareGraphicsExtensions{.eps}
\fi
% graphicx was written by David Carlisle and Sebastian Rahtz. It is
% required if you want graphics, photos, etc. graphicx.sty is already
% installed on most LaTeX systems. The latest version and documentation
% can be obtained at:
% http://www.ctan.org/pkg/graphicx
% Another good source of documentation is "Using Imported Graphics in
% LaTeX2e" by Keith Reckdahl which can be found at:
% http://www.ctan.org/pkg/epslatex
%
% latex, and pdflatex in dvi mode, support graphics in encapsulated
% postscript (.eps) format. pdflatex in pdf mode supports graphics
% in .pdf, .jpeg, .png and .mps (metapost) formats. Users should ensure
% that all non-photo figures use a vector format (.eps, .pdf, .mps) and
% not a bitmapped formats (.jpeg, .png). The IEEE frowns on bitmapped formats
% which can result in "jaggedy"/blurry rendering of lines and letters as
% well as large increases in file sizes.
%
% You can find documentation about the pdfTeX application at:
% http://www.tug.org/applications/pdftex
% *** MATH PACKAGES ***
%
%\usepackage{amsmath}
% A popular package from the American Mathematical Society that provides
% many useful and powerful commands for dealing with mathematics.
%
% Note that the amsmath package sets \interdisplaylinepenalty to 10000
% thus preventing page breaks from occurring within multiline equations. Use:
%\interdisplaylinepenalty=2500
% after loading amsmath to restore such page breaks as IEEEtran.cls normally
% does. amsmath.sty is already installed on most LaTeX systems. The latest
% version and documentation can be obtained at:
% http://www.ctan.org/pkg/amsmath
% *** SPECIALIZED LIST PACKAGES ***
%
%\usepackage{algorithmic}
% algorithmic.sty was written by Peter Williams and Rogerio Brito.
% This package provides an algorithmic environment fo describing algorithms.
% You can use the algorithmic environment in-text or within a figure
% environment to provide for a floating algorithm. Do NOT use the algorithm
% floating environment provided by algorithm.sty (by the same authors) or
% algorithm2e.sty (by Christophe Fiorio) as the IEEE does not use dedicated
% algorithm float types and packages that provide these will not provide
% correct IEEE style captions. The latest version and documentation of
% algorithmic.sty can be obtained at:
% http://www.ctan.org/pkg/algorithms
% Also of interest may be the (relatively newer and more customizable)
% algorithmicx.sty package by Szasz Janos:
% http://www.ctan.org/pkg/algorithmicx
% *** ALIGNMENT PACKAGES ***
%
%\usepackage{array}
% Frank Mittelbach's and David Carlisle's array.sty patches and improves
% the standard LaTeX2e array and tabular environments to provide better
% appearance and additional user controls. As the default LaTeX2e table
% generation code is lacking to the point of almost being broken with
% respect to the quality of the end results, all users are strongly
% advised to use an enhanced (at the very least that provided by array.sty)
% set of table tools. array.sty is already installed on most systems. The
% latest version and documentation can be obtained at:
% http://www.ctan.org/pkg/array
% IEEEtran contains the IEEEeqnarray family of commands that can be used to
% generate multiline equations as well as matrices, tables, etc., of high
% quality.
% *** SUBFIGURE PACKAGES ***
%\ifCLASSOPTIONcompsoc
% \usepackage[caption=false,font=footnotesize,labelfont=sf,textfont=sf]{subfig}
%\else
% \usepackage[caption=false,font=footnotesize]{subfig}
%\fi
% subfig.sty, written by Steven Douglas Cochran, is the modern replacement
% for subfigure.sty, the latter of which is no longer maintained and is
% incompatible with some LaTeX packages including fixltx2e. However,
% subfig.sty requires and automatically loads Axel Sommerfeldt's caption.sty
% which will override IEEEtran.cls' handling of captions and this will result
% in non-IEEE style figure/table captions. To prevent this problem, be sure
% and invoke subfig.sty's "caption=false" package option (available since
% subfig.sty version 1.3, 2005/06/28) as this is will preserve IEEEtran.cls
% handling of captions.
% Note that the Computer Society format requires a sans serif font rather
% than the serif font used in traditional IEEE formatting and thus the need
% to invoke different subfig.sty package options depending on whether
% compsoc mode has been enabled.
%
% The latest version and documentation of subfig.sty can be obtained at:
% http://www.ctan.org/pkg/subfig
% *** FLOAT PACKAGES ***
%
%\usepackage{fixltx2e}
% fixltx2e, the successor to the earlier fix2col.sty, was written by
% Frank Mittelbach and David Carlisle. This package corrects a few problems
% in the LaTeX2e kernel, the most notable of which is that in current
% LaTeX2e releases, the ordering of single and double column floats is not
% guaranteed to be preserved. Thus, an unpatched LaTeX2e can allow a
% single column figure to be placed prior to an earlier double column
% figure.
% Be aware that LaTeX2e kernels dated 2015 and later have fixltx2e.sty's
% corrections already built into the system in which case a warning will
% be issued if an attempt is made to load fixltx2e.sty as it is no longer
% needed.
% The latest version and documentation can be found at:
% http://www.ctan.org/pkg/fixltx2e
%\usepackage{stfloats}
% stfloats.sty was written by Sigitas Tolusis. This package gives LaTeX2e
% the ability to do double column floats at the bottom of the page as well
% as the top. (e.g., "\begin{figure*}[!b]" is not normally possible in
% LaTeX2e). It also provides a command:
%\fnbelowfloat
% to enable the placement of footnotes below bottom floats (the standard
% LaTeX2e kernel puts them above bottom floats). This is an invasive package
% which rewrites many portions of the LaTeX2e float routines. It may not work
% with other packages that modify the LaTeX2e float routines. The latest
% version and documentation can be obtained at:
% http://www.ctan.org/pkg/stfloats
% Do not use the stfloats baselinefloat ability as the IEEE does not allow
% \baselineskip to stretch. Authors submitting work to the IEEE should note
% that the IEEE rarely uses double column equations and that authors should try
% to avoid such use. Do not be tempted to use the cuted.sty or midfloat.sty
% packages (also by Sigitas Tolusis) as the IEEE does not format its papers in
% such ways.
% Do not attempt to use stfloats with fixltx2e as they are incompatible.
% Instead, use Morten Hogholm'a dblfloatfix which combines the features
% of both fixltx2e and stfloats:
%
% \usepackage{dblfloatfix}
% The latest version can be found at:
% http://www.ctan.org/pkg/dblfloatfix
% *** PDF, URL AND HYPERLINK PACKAGES ***
%
%\usepackage{url}
% url.sty was written by Donald Arseneau. It provides better support for
% handling and breaking URLs. url.sty is already installed on most LaTeX
% systems. The latest version and documentation can be obtained at:
% http://www.ctan.org/pkg/url
% Basically, \url{my_url_here}.
% *** Do not adjust lengths that control margins, column widths, etc. ***
% *** Do not use packages that alter fonts (such as pslatex). ***
% There should be no need to do such things with IEEEtran.cls V1.6 and later.
% (Unless specifically asked to do so by the journal or conference you plan
% to submit to, of course. )
% correct bad hyphenation here
\hyphenation{op-tical net-works semi-conduc-tor}
\newcommand{\citep}[1]{\cite{#1}}
% \usepackage[numbers]{natbib}
\usepackage{url}
\usepackage{booktabs}
\usepackage{amsfonts}
\usepackage{amsmath}
\usepackage{amssymb}
\usepackage{graphicx}
\newcommand{\cescore}[1]{\mathrm{CE}^{\mathrm{score}}_{#1}}
\newcommand{\ceorder}[1]{\mathrm{CE}^{\mathrm{order}}_{#1}}
\newcommand{\cepos}[1]{\mathrm{CE}^{\mathrm{pos}}_{#1}}
\begin{document}
%
% paper title
% Titles are generally capitalized except for words such as a, an, and, as,
% at, but, by, for, in, nor, of, on, or, the, to and up, which are usually
% not capitalized unless they are the first or last word of the title.
% Linebreaks \\ can be used within to get better formatting as desired.
% Do not put math or special symbols in the title.
\title{Plausible Counterfactual Explanations \\of Recommendations}
% author names and affiliations
% use a multiple column layout for up to three different
% affiliations
\author{\IEEEauthorblockN{Jakub Černý}
\IEEEauthorblockA{Dateio\\
Prague, Czechia\\
Email: jakub.cerny@dateio.eu}
\and
\IEEEauthorblockN{Jiří Němeček}
\IEEEauthorblockA{Faculty of Electrical Engineering\\
Czech Technical University in Prague\\
Prague, Czechia\\
Email: contact@nemecekjiri.cz}
\\
\IEEEauthorblockN{Jakub Mareček}
\IEEEauthorblockA{Faculty of Electrical Engineering\\
Czech Technical University in Prague\\
Prague, Czechia\\
Email: jakub.marecek@fel.cvut.cz}
\and
\IEEEauthorblockN{Ivan Dovica}
\IEEEauthorblockA{Dateio\\
Prague, Czechia\\
Email: ivan.dovica@dateio.eu}
}
% conference papers do not typically use \thanks and this command
% is locked out in conference mode. If really needed, such as for
% the acknowledgment of grants, issue a \IEEEoverridecommandlockouts
% after \documentclass
% for over three affiliations, or if they all won't fit within the width
% of the page (and note that there is less available width in this regard for
% compsoc conferences compared to traditional conferences), use this
% alternative format:
%
%\author{\IEEEauthorblockN{Michael Shell\IEEEauthorrefmark{1},
%Homer Simpson\IEEEauthorrefmark{2},
%James Kirk\IEEEauthorrefmark{3},
%Montgomery Scott\IEEEauthorrefmark{3} and
%Eldon Tyrell\IEEEauthorrefmark{4}}
%\IEEEauthorblockA{\IEEEauthorrefmark{1}School of Electrical and Computer Engineering\\
%Georgia Institute of Technology,
%Atlanta, Georgia 30332--0250\\ Email: see http://www.michaelshell.org/contact.html}
%\IEEEauthorblockA{\IEEEauthorrefmark{2}Twentieth Century Fox, Springfield, USA\\
%Email: homer@thesimpsons.com}
%\IEEEauthorblockA{\IEEEauthorrefmark{3}Starfleet Academy, San Francisco, California 96678-2391\\
%Telephone: (800) 555--1212, Fax: (888) 555--1212}
%\IEEEauthorblockA{\IEEEauthorrefmark{4}Tyrell Inc., 123 Replicant Street, Los Angeles, California 90210--4321}}
% use for special paper notices
%\IEEEspecialpapernotice{(Invited Paper)}
% make the title area
\maketitle
% As a general rule, do not put math, special symbols or citations
% in the abstract
\begin{abstract}
Explanations play a variety of roles in various recommender systems, from a legally mandated afterthought, through an integral element of user experience, to a key to persuasiveness.
A natural and useful form of an explanation is the Counterfactual Explanation (CE). We present a method for generating highly plausible CEs in recommender systems and evaluate it both numerically and with a user study.
\end{abstract}
% no keywords
% For peer review papers, you can put extra information on the cover
% page as needed:
% \ifCLASSOPTIONpeerreview
% \begin{center} \bfseries EDICS Category: 3-BBND \end{center}
% \fi
%
% For peerreview papers, this IEEEtran command inserts a page break and
% creates the second title. It will be ignored for other modes.
\IEEEpeerreviewmaketitle
\section{Introduction}
Explanations in recommender systems (RS) have been repeatedly shown to improve user experience \citep{herlocker_explaining_2000,shang_why_2022} by enhancing trust in the RS, user satisfaction with the RS, and transparency of the system \citep{nunes_systematic_2017}. Recommender explanations (or ``recsplanations'' \citep{mcinerney_explore_2018}) also measurably increase the persuasiveness of the RS \citep{nunes_systematic_2017}
and help users make decisions faster.
Shang et al. \citep{shang_why_2022} showed that around half of users (up to 62\% in e-commerce) are \textit{highly interested} in obtaining an explanation.
In addition, regulation \citep{gdpr,CCPA,ai-act} requires explanations for automated decisions made using personal data.
There are many forms of explanations \citep{ge_survey_2024} of RS. Here, we focus on Counterfactual Explanations (CEs, \citep{wachter_counterfactual_2017}) in particular.
% what are they?
Counterfactual explanations, in general, answer the question ``How would the input have to change for the output to change a given way?'' For example: ``In order to recommend Kill Bill, your rating of Pulp Fiction would have to be at least 4 stars out of 5.''
Counterfactual explanations are well understandable and considered useful by users \citep{shang_why_2022} and even suggested\footnote{A direct citation: ``[...] it is sufficiently transparent and intelligible to inform the data subject of the extent to which a variation in the personal data taken into account would have led to a different result.''} by recent EU court judgment
\citep{gdpr_supports_ce_case}. This makes them well-suited for practical use.
% \jirimargincomment{neco z toho why not paperu, consider taky From intrinsic to counterfactual: On the explainability of contextualized recommender systems}
Counterfactual explanations are an active area of research outside of recommender systems \citep{guidotti_counterfactual_2022} but also within the RS domain \citep{ge_survey_2024, baklanov_ceers_2024}.
% However, one crucial aspect is somewhat omitted in the literature on CEs for recommenders. This aspect
One may wish for the CE to be actionable (e.g., not recommending becoming younger or less educated) and sparse (i.e., utilizing as few changes as possible) and these qualities remain the same across a variety of application domains. However, there is an overlooked desideratum of \textit{plausibility}, which is rarely addressed by CE methods for recommenders. In agreement with the commonly considered definition \citep{guidotti_counterfactual_2022}, we define plausibility as the likelihood of the counterfactual with respect to the training data.
Let us consider the same Tarantino example as above, with the current user having rated Pulp Fiction 4 stars and asking about the recommendation of Kill Bill. A counterfactual ``Kill Bill would not be recommended, if you rated Pulp Fiction 3 stars.'' is more plausible than ``Kill Bill would not be recommended, if you rated Pulp Fiction 1 star'', because it is more likely, conditional on the attribute values of the present user of the RS.
While both counterfactuals are perfectly valid, %but assuming that few similar users with such an out-of-distribution rating exist, this is unlikely to be applicable. A
a user then might consider the latter explanation less useful, since such a low rating is outside of the margin for error in the ratings, due to mood at the time of rating or watching the movie or similar.
%\jirimargincomment{an actual example would be better here}
There are several general-purpose methods for the generation of CE with plausibility objectives. We build on LiCE \citep{nemecek2025generating}, which has recently been shown to outperform other comparable methods.
Our contributions are as follows:
\begin{itemize}
\item We propose a modification of LiCE as a method for plausible Counterfactual Explanations of recommendations.
% \item We specify of
\item We compare various setups on the plausibility of CEs.
\item We validate the proposed modification with a user study.
\end{itemize}
With our contributions, we open the question of plausible explanations of Recommender Systems, evaluate the top-performing model, and describe the pros and cons of each setup. Our work enables practitioners to make informed decisions in order to improve both the user experience and the performance of their system \citep{nunes_systematic_2017}.
% ------------- outline ---------------
% - explanations are useful
% - in multiple dimensions (persuasiveness etc.)
% - well-studied in general postponed to related works
% - counterfactuals are great (define them)
% - supported by regulation (link to the court decision)
% - studied little in this domain? NOPE
% - quite popular, (recent surveys)
% - special focus on plausibility
% - similar arguments as in lice [FIND an example where the plausible CE is clearly better otherwise make one up]
% - there are many methods, but the recent best, with support for binary vars and tweakable to one's liking, is LiCE
% - we evaluate it [on more data???? and more methods????], considering a user study
% \jiricomment{our contributions - listed clearly - explain why relevant for RecSys people...}
\section{Preliminaries}
% \paragraph{Notation}
% We consider the following problem.
Let there be $D$ items. Let $x_i \in \mathcal{X}$ be the vector of interactions (e.g., clicks, views, ratings...) of a user $i$, where $\mathcal{X} \subseteq \mathbb{R}^D$ is the $D$-dimensional space of such interactions. Each user-item interaction is a scalar that can take real values (e.g., length of an interaction), but they are oftentimes integer or binary (e.g., item ratings or purchases). We have a set $\mathcal{D} = \{x_1, x_2, \ldots , x_N\}$ of $N$ such vectors, representing $N$ users. We refer to this set as the (training) dataset.
\paragraph{LiCE}
% with support for binary vars and tweakable to one's liking, is LiCE
We modify LiCE \citep{nemecek2025generating}, a method for generating Likely Counterfactual Explanations (CEs). LiCE utilizes on Mixed-Integer Optimization (MIO, \cite{wolsey_integer_2021}), a popular framework for solving NP-Hard optimization problems \cite{kochProgressMathematicalProgramming2022}, and tractable probabilistic models \cite{poon_sum-product_2011,NEURIPS2023_6b61c278,sidheekh2023probabilistic}, a popular framework for modelling uncertainty, somewhat resembling neural networks in the compositionality of the model. In particular, LiCE utilizes a Sum-Product Network (SPN, \cite{poon_sum-product_2011}) to estimate the likelihood of a counterfactual. SPNs and MIO are particularly suitable for the use in RS, where the input variables are often discrete, which MIO and SPNs model naturally.
\paragraph{SPN}
To elaborate: SPNs are a class of tractable probabilistic models, taking the form of a directed acyclic graph (DAG), where each node represents a distribution over a defined set of variables (item interactions, in our case). The leaf nodes in the DAG represent simple distributions, often univariate. The non-leaf nodes are of one of two types. There are sum nodes, which produce a weighted mixture of the distributions in the ``successor nodes''. The remaining non-leaf nodes are product nodes, which produce the product of distributions in the ``successor nodes''.
% However, they mostly ommit Counterfactual Explanations (CEs), with the exception of the work by Shang et al. \citep{shang_why_2022}, which does not propose a method for generating CEs, but rather validates the usefulness and user preference of counterfactuals.
\section{Related works}
A quarter century of work on explaining recommendations since the pioneering work by Herlocker et al. \citep{herlocker_explaining_2000} addressed the question of explaining black-box recommenders. The 25 years of development and the diverse landscape of explaining recommender systems is described by a multitude of reviews \citep{mcsherry_explanation_2005,nunes_systematic_2017}, taxonomies \citep{ friedrich_taxonomy_2011,nunes_systematic_2017}, surveys \citep{vultureanu-albisi_survey_2022,wardatzky_evaluating_2024,ge_survey_2024}.
%even PhD theses \citep{melchiorre_explainable_2024}.
We shall focus on the Counterfactual Explanations, best surveyed recently by Ge et al. \citep{ge_survey_2024}, and well motivated by Shang et al. \citep{shang_why_2022}, based on user studies and interviews.
\paragraph{Generating Counterfactual Explanations in Recommender Systems.} Let us mention several key methods: PRINCE \citep{ghazimatin_prince_2020}, focused on finding CEs over a heterogeneous information network, assuming that personalized PageRank is used to infer recommendations.
%
ACCENT \citep{tran_counterfactual_2021} provides CEs for recommenders based on neural networks, but it only considers a limited number of items and seeks a change of the top-ranked one, rather than considering the entire ranking.
% taky maj nic moc fidelity?
%
CR-VAE \citep{xu_learning_2021} utilizes a Variational Auto-Encoder (VAE) to obtain CEs and then finds causal dependencies.
% co je s timhle spatne, to nevim :D ~97% validity (asi), taky plausible... maybe shit proximity??? - jestli rozumim ~20 edits? - nope, in latent space - would have to do my own?
%
PGCE \citep{he_prototype-guided_2023} also utilizes VAE. It uses gradient-based methods, with an additional loss term on the distance to a prototype (mean of almost recommended items) in the latent space. The authors evaluate ``realism'', but they measure it as normalized $\ell_2$-distance from the factual, while we propose utilizing a probabilistic model.
%
CountER \citep{tan_counterfactual_2021} focuses on the items being recommended (rather than user-item interactions), proposing similar counterfactual items that would no longer be recommended.
%
CF$^2$ \citep{Tan_2022} and GREASE \citep{chen_grease_2022} are methods of generating CEs for recommenders based on GNNs.
%
Finally, Kaffes et al. \citep{kaffes_model-agnostic_2021} generate model-agnostic CEs by search, and Zhong and Negre \citep{zhong_shap-enhanced_2022} expand on that by ordering the interactions used in the search according to their SHAP values.
We refer to \cite{ge2024survey} for a recent survey.
\paragraph{The contrast.}
Our approach improves on the state of the art by reliably finding the closest counterfactual explanations for most recommender models without the need of exhaustive search. The generated CEs have a high likelihood with respect to the data distribution and are faithful to the model.
% \jirimargincomment{are others less plausible? we'll never know}
We obtain these by modifying LiCE \citep{nemecek2025generating}, a well-performing method for generating plausible counterfactuals and applying it to recommender systems. Our modification is also openly accessible at \url{https://github.com/Epanemu/LiCE4Recommenders}.
We know of no other publicly implemented method that constrains CEs by removing an item from top-$k$ recommendations, which is what we find desirable.
% \jirimargincomment{briefly discuss: MIO - in recommender systems?, SPNs for recmmender systems?
% show that plausibility is omitted...
% }
\section{Our Method}
% \jirimargincomment{check if the other models next to ease are also representable using MIO - discuss that here through "best guess" claims.}
To model the counterfactual generation for ranking as a Mixed-Integer Optimization (MIO) problem, we have to formulate the ranking process to ensure validity. Throughout, we assume the availability of a scoring function $f_j(x_i)$ for each item $j$ and select the top-$k$ items with the highest values to be recommended.
This holds for Neural Collaborative Filtering (NCF, \citep{he_neural_2017}), EASE \cite{steck_embarrassingly_2019}, and even more recent VASP \citep{vancura_deep_2021}, for example.
% Given that EASE is just a matrix multiplication, formulating the prediction score $y_i$ of a merchant $i$ is
Subsequently, the MIO formulation reads
\begin{align}
\min & \sum_{j=1}^{D} |x_{i,j} - x_{i,j}^{\prime}| \label{eq:objective} \\
% \text{such that} \; & y_j = B_{j,:} \cdot x^{\prime} & \forall j \in \{1, \ldots, n\} \\
\text{such that} \; & y_j = f_j(x_i^{\prime}) & \forall j \in \{1, \ldots, D\} \\
& x^{\prime}_{i,j} \in \{0, 1\}, y_j \in \mathbb{R} & \forall j \in \{1, \ldots, D\},
\end{align}
where $x_i$ is the factual (the input vector of user $i$) and $x_i^{\prime}$ is the counterfactual. The absolute value can be readily linearized. In the binary case, $x_j - x_{i,j}^{\prime}$ if $x_{i,j} = 1$ and $x_{i,j}^{\prime} - x_{i,j}$ if $x_{i,j} = 0$. In the continuous case, one can introduce \cite{wolsey_integer_2021,nemecek2025generating} auxiliary variable. Note that we slightly diverge from LiCE in using a plain $\ell_1$ distance. This is because we mainly work with binary variables.
We must also be able to formulate the scoring function $f_j$ within MIO. For Neural models \citep{he_neural_2017} including VAEs \citep{liang_variational_2018}, we refer to the widely studied topic of representing Neural Networks in MIO (e.g., \citep{anderson_strong_2020}), where ready-made implementations exist \citep{bynum_pyomo_2021}.
For EASE \citep{steck_embarrassingly_2019} (and similarly for SLIM \citep{ning_slim_2011}), $f_j(x^{\prime})$ is simply $B_{j,:} \cdot x^{\prime}$, where $B_{j,:}$ is the $j$-th row of the matrix $B$.
Even more recent high-performing models, such as H+ Vamp Gated \citep{kim_enhancing_2019} or VASP
% a combination of multiple Neural architectures
\citep{vancura_deep_2021} could be formulated, using approximations or non-linearity in the formulation.
To avoid the trivial optimum $x_i^{\prime} := x_i$, we constrain the value $y_c$ of the item $c$, for which we wish to find the explanation.
Naively, we could define a counterfactual $\cescore{\tau}$ with the minimal change required for the score of merchant $i$ to drop below a threshold $\tau$:
\begin{equation}
y_i \le \tau.
\end{equation}
Unfortunately, this does not ensure that item $c$ is not be recommended, given the counterfactual. It might happen that the counterfactual change would decrease the scores of all items, and the ordering would remain the same. To establish what makes item $c$ rank high, we define a counterfactual $\cepos{\rho}$, which is the minimal change in $x_i$, such that the item $c$ is among the top-$\rho$ items.
In the MIO formulation, we introduce $D-1$ binary variables $r_j$, which equal 1 if and only if the merchant $j$ ranks above the merchant $i$. Formulating the constraints
\begin{align}
y_c & \le y_j + (1 - r_j) * M & \forall j \in \{1,\ldots,D\} \setminus \{c\} \\
\sum_{j = 1}^D r_j & \ge \rho, \label{eq:rank}
\end{align}
where $M$ is a ``big-M'' constant \citep{wolsey_integer_2021} large enough to cover the span of possible $y_c$ values. It can be computed for each $j$ by subtracting the minimal value of $y_j$ from the maximal possible value of $y_c$. We want the big-M to be as low as possible, since a high $M$ can cause numerical issues. The constraint \eqref{eq:rank} ensures that $\rho$ items have a score greater than or equal to that of the item $c$. This corresponds to a rank of at least $\rho + 1$.
% Finally, we might want to know how we could achieve some specific ordering of a subset of $m$ items.
% % For example, we wish to know why an item $i$ is ranked higher than comparable items $j$ and $k$. We thus ask what would need to happen so that the merchant $j$ was first, then $k$ and the merchant $i$ was last.
% % \jirimargincomment{make up a proper example usecase or just remove this}
% We define a counterfactual $\ceorder{O}$, requiring that merchants follow the ordering $O = (o_1, \ldots o_m)$, meaning that $y_{o_l} \ge y_{o_{l+1}}$ for all $l \in \{1, \ldots, m-1\}$.
% % In the previous example, $O = (j, k, i)$.
% And this is precisely how we formulate the constraints.
% We formulate this as
% \begin{equation}
% y_{o_l} \ge y_{o_{l+1}} \quad \forall l \in \{1, \ldots, m-1\}.
% \end{equation}
\subsection{SPN variants}
By default, LiCE utilizes SPNs modeled on all features, i.e., considers each item separately. While having the best granularity, one might need a lot of data to capture the joint distribution without overfitting. It might also lead to SPNs, which are costly to optimize over. Luckily, items often belong to some categories or come with tags. Let there be $K$ categories, each described by a set $\mathcal{K}_j$, which contains items belonging to the category $j$. We can then aggregate the sparse information into $K$-dimensional feature vectors and use those for training.
Denote the resulting vector of aggregated values for user $i$ by $z_i$. We evaluate three aggregating functions. The first aggregating function is the sum, which can be formulated in MIO directly by
\begin{equation}
z_{i,j} = \sum_{k \in \mathcal{K}_j} x_{i, k}^{\prime} \quad \forall j \in \{1, 2, \ldots, K\}.
\end{equation}
The second aggregating function is a mean, which is modeled the same way, but multiplied by the inverse of the number of items in a given category
\begin{equation}
z_{i,j} = \frac{1}{|\mathcal{K}_j|} \sum_{k \in \mathcal{K}_j} x_{i, k}^{\prime} \quad \forall j \in \{1, 2, \ldots, K\}.
\end{equation}
Utilizing the mean is fundamentally different from the sum in that the sum retains the discrete property of input data, while the mean leads to non-discrete random variables, which are handled differently in the SPN.
Finally, there is the disjunction (logical or) modeled by
\begin{align}
z_{i,j} \le & \sum_{k \in \mathcal{K}_j} x_{i, k}^{\prime} & \forall j \in \{1, 2, \ldots , K\} \label{eq:force_0} \\
z_{i,j} \ge & \frac{1}{|\mathcal{K}_j|} \sum_{k \in \mathcal{K}_j} x_{i, k}^{\prime} & \forall j \in \{1, 2, \ldots , K\} \label{eq:force_1} \\
z_{i,j} \in & \{0, 1\} & \forall j \in \{1, 2, \ldots , K\},
\end{align}
where we assume all $x_{i,k}^{\prime}$ are binary, the constraint \eqref{eq:force_0} forces the binary $z_{i,j}$ to be zero when all items in a given category are not interacted with (have value 0). In all other cases, the constraint \eqref{eq:force_1} will force $z_{i,j}$ to equal one. This formulation can be extended to non-binary inputs by introducing appropriate scaling factors.
\section{Numerical Evaluation}
% \subsection{Setup}
\begin{table*}[t]
\centering
\caption{Plausible CE within a time limit: proportion of
% globally optimal
results obtained within the time limit.
Only for 8.6\% of CEs, their global optimality could not be proven within the time limit of
10 minutes. Median run-time was 11 seconds.}
\resizebox{\textwidth}{!}{\begin{tabular}{llcccccccccc}
\toprule
% \multicolumn{2}{l}{Variants}
& & \multicolumn{4}{c}{Optimize ($\alpha = 0.1$)} & \multicolumn{4}{c}{Threshold (median LL)} & \multicolumn{2}{c}{No SPN} \\
\cmidrule(lr){3-6} \cmidrule(lr){7-10} \cmidrule{11-12}
% \multicolumn{2}{l}{Input formats}
& & \multicolumn{3}{c}{Binarized input} & Rating & \multicolumn{3}{c}{Binarized input} & Rating & Binarized input & Rating \\
\cmidrule(lr){3-5} \cmidrule(lr){6-6} \cmidrule(lr){7-9} \cmidrule(lr){10-10} \cmidrule{11-11} \cmidrule(lr){12-12}
% \multicolumn{2}{l}{Aggregator functions}
& Validity variant & Disjunction & Sum & Mean & Mean & Disjunction & Sum & Mean & Mean & N/A & N/A \\
\cmidrule(lr){1-2}
\cmidrule(lr){3-12}
% \cmidrule(lr){3-6} \cmidrule(lr){7-10} \cmidrule(lr){11-12}
Amazon & $k$-th rank & 1.00 & 0.99 & 0.97 & 0.97 & 0.92 & 1.00 & 0.00 & 0.00 & 1.00 & 1.00 \\
& score threshold & 1.00 & 0.95 & 1.00 & 1.00 & 0.94 & 0.95 & 0.00 & 0.00 & 1.00 & 1.00 \\
Yelp & $k$-th rank & 1.00 & 1.00 & 0.30 & 0.07 & 0.86 & 0.00 & 0.00 & 0.00 & 1.00 & 1.00 \\
& score threshold & 1.00 & 1.00 & 0.95 & 0.88 & 1.00 & 0.00 & 0.00 & 0.00 & 1.00 & 1.00 \\
Netflix & $k$-th rank & 1.00 & 1.00 & 0.70 & 0.51 & 0.99 & 0.00 & 0.00 & 0.33 & 1.00 & 1.00 \\
& score threshold & 1.00 & 1.00 & 0.89 & 0.62 & 1.00 & 0.00 & 0.00 & 0.33 & 1.00 & 1.00 \\
% \midrule
% Dateio - categories & $k$-th rank & 1.00 & 1.00 & 1.00 & - & 0.58 & 1.00 & 1.00 & - & 1.00 & - \\
% & score threshold & 1.00 & 1.00 & 1.00 & - & 0.58 & 1.00 & 1.00 & - & 1.00 & - \\
% Dateio - tags & $k$-th rank & 1.00 & 1.00 & 1.00 & - & 0.30 & 1.00 & 0.00 & - & 1.00 & - \\
% & score threshold & 1.00 & 1.00 & 1.00 & - & 0.12 & 1.00 & 0.00 & - & 1.00 & - \\
\bottomrule
\end{tabular}}
\label{tab:rates}
\end{table*}
\begin{table*}[t]
\centering
\caption{A tradeoff between plausibility and similarity
evaluated as log-likelihood and $\ell_1$ distance: mean and standard deviation over all retrieved CEs of a given configuration. %Following results from Table \ref{tab:rates}, we do not show results of sum and mean aggregator functions for binarized input results and rating results for thresholding variant, due to low success rates. These results suggest a tradeoff between plausibility and similarity.
}
\resizebox{\textwidth}{!}{\begin{tabular}{llcccccccccc}
\toprule
& & \multicolumn{5}{c}{\textbf{Plausibility (Log-Likelihood)}} & \multicolumn{5}{c}{\textbf{Similarity ($\ell_1$ Distance)}} \\
\cmidrule(lr){3-7} \cmidrule(lr){8-12}
& & \multicolumn{3}{c}{Binarized input - Disjunction} & \multicolumn{2}{c}{Rating - Mean} & \multicolumn{3}{c}{Binarized input - Disjunction} & \multicolumn{2}{c}{Rating - Mean} \\
\cmidrule(lr){3-5} \cmidrule(lr){6-7} \cmidrule(lr){8-10} \cmidrule(lr){11-12}
& & Optimize & Threshold & No SPN & Optimize & No SPN & Optimize & Threshold & No SPN & Optimize & No SPN \\
\cmidrule(lr){1-7} \cmidrule(lr){8-12}
Amazon & $k$-th & $-18.44 \pm 12.23$ & $-12.08 \pm 3.16$ & $-19.57 \pm 13.45$ & $1695 \pm 31$ & $1693 \pm 34$ & $1.16 \pm 0.49$ & $2.60 \pm 3.24$ & $1.08 \pm 0.43$ & $2.11 \pm 2.16$ & $0.66 \pm 0.38$ \\
& score & $-18.40 \pm 11.93$ & $-12.21 \pm 3.42$ & $-19.61 \pm 13.42$ & $1695 \pm 35$ & $1694 \pm 34$ & $1.19 \pm 0.46$ & $2.82 \pm 3.94$ & $1.11 \pm 0.37$ & $2.33 \pm 3.19$ & $0.54 \pm 0.37$ \\
Yelp & $k$-th & $-89.11 \pm 38.68$ & $-78.33 \pm 13.72$ & $-103.21 \pm 46.04$ & $3185 \pm 113$ & $3122 \pm 148$ & $2.71 \pm 2.04$ & $4.16 \pm 7.69$ & $1.65 \pm 0.98$ & $2.54 \pm 1.98$ & $1.07 \pm 0.86$ \\
& score & $-89.23 \pm 39.80$ & $-82.42 \pm 14.69$ & $-103.06 \pm 45.99$ & $3184 \pm 84$ & $3122 \pm 148$ & $2.44 \pm 1.57$ & $5.02 \pm 10.89$ & $1.62 \pm 0.89$ & $6.92 \pm 5.17$ & $1.01 \pm 0.79$ \\
Netflix & $k$-th & $-23.44 \pm 9.10$ & $-19.67 \pm 2.89$ & $-23.61 \pm 9.25$ & $381 \pm 44$ & $367 \pm 60$ & $0.71 \pm 1.03$ & $4.07 \pm 8.34$ & $0.69 \pm 0.90$ & $0.53 \pm 1.13$ & $0.39 \pm 0.73$ \\
& score & $-23.33 \pm 9.11$ & $-19.66 \pm 2.86$ & $-23.62 \pm 9.18$ & $385 \pm 45$ & $368 \pm 60$ & $2.07 \pm 1.37$ & $4.87 \pm 5.57$ & $2.07 \pm 1.37$ & $1.84 \pm 4.01$ & $1.28 \pm 1.12$ \\
% \cmidrule(lr){1-7} \cmidrule(lr){8-12}
% Dateio - categs & nth & $-7.91 \pm 3.03$ & $-5.91 \pm 0.64$ & $-7.93 \pm 3.02$ & - & - & $1.00 \pm 0.00$ & $1.00 \pm 0.00$ & $1.00 \pm 0.00$ & - & - \\
% & score & $-7.91 \pm 3.03$ & $-5.91 \pm 0.64$ & $-7.92 \pm 3.03$ & - & - & $1.00 \pm 0.00$ & $1.00 \pm 0.00$ & $1.00 \pm 0.00$ & - & - \\
% Dateio - tags & nth & $-310.5 \pm 56.6$ & $-83.40 \pm 0.39$ & $-372.8 \pm 70.2$ & - & - & $4.96 \pm 1.87$ & $87.02 \pm 18.69$ & $1.00 \pm 0.00$ & - & - \\
% & score & $-302.7 \pm 59.5$ & $-83.64 \pm 0.04$ & $-366.9 \pm 61.9$ & - & - & $4.75 \pm 2.17$ & $92.17 \pm 15.33$ & $1.00 \pm 0.00$ & - & - \\
\bottomrule
\end{tabular}}
\label{tab:stats}
\end{table*}
We evaluate our method on three public datasets: Yelp \citep{yelp_data}, Amazon book reviews \citep{amazon_review_data}, and Netflix competition \citep{netflix_data}. In addition to that, we test on a private Dateio dataset of transactions. There, one predicts at which merchant a bank's customer is likely to shop, in order to recommend suitable cashback offers to encourage spending.
% (We refer to all datasets by the company name; the private one being called Dateio.)
All of the public datasets contain ratings on a scale of 1-5. We normalize these values to the $[0,1]$ range. Additionally, we test on data where ratings are binarized by mapping all non-zeros to 1. Binarized Yelp data are then very similar to the Dateio data, given that items are sellers.
We prune users and items with few interactions (reviews/purchases).
To model the data distribution, working with the full input space would be inefficient. We therefore group items into the categories provided with the data, and in the case of Netflix, we group movies by their release year. In Dateio, each merchant is assigned a single category and some (more granular) tags. For example, a category might be Food \& Drink, and a tag could be American Fastfood. Values of items belonging to each group are aggregated using one of three functions: mean, sum, or disjunction. We test sum and disjunction only on the binarized variant of the data, since those functions retain the discreteness of the data, which SPNs can utilize.
% , see previous sections.
% \jiricomment{add comments on the various setups -
% mean dis sum spns,
% rating/binary,
% median optimize nospn,
% score nth,
% topk
% we only add the decrease of the score, no increase
% we only test for EASE -
% $f_j(x_i) = B_{j,:} \cdot x_i$ where $B_{j,:}$ is the $j$-th row of the weight matrix.
% we use highspy - OSS
% }
In addition to the SPN variant, we also choose how we utilize the SPN output. Following \citep{nemecek2025generating}, we test 3 variants. In the first variant, the SPN is omitted. In the other two, we either impose a constraint on the CE likelihood to be at least equal to the median likelihood of the training data, or we subtract the likelihood into the objective \eqref{eq:objective} multiplied by $\alpha = 0.1$, effectively minimizing a linear combination of $\ell_1$ distance and negative log-likelihood.
We take 50 users, and retrieve the top $k$ recommended items for each of them. Then, for each of the top three items, we generate a counterfactual input, such that the item would rank outside of the top $k$ items of a given user. We choose $k$ to be 5 or 10, which roughly corresponds to the item not being shown on a single row of recommendations, or on the first page of results.
We test the variant where the drop is guaranteed (called $k$-th) and a variant where we constrain the item's score by the original score of the $k$-th element (called score).
In terms of allowed change, we only allow the interaction values to decrease, since we strive for explanations in the form of ``you have been recommended X because you interacted with Y,'' which translates to a counterfactual where item Y is removed from the input vector. For example, recommending a movie because a person had \textit{not} seen some other movie would be non-obvious.
For the three public datasets, we do 3-fold cross-validation, training 3 recommenders for each dataset, in order for the evaluation to be less dependent on the RS performance \citep{mohammadi_are_2024}.
% For the Dateio dataset, we utilize only the production recommender as a baseline.
Throughout, we utilize EASE \citep{steck_embarrassingly_2019} in all of the data splits.
We model the formulation using the Pyomo library \citep{bynum_pyomo_2021} and solve it using an open-source solver, HiGHS \citep{huangfuParallelizingDualRevised2018}. We optimize each formulation for up to 10 minutes (a conservative limit; the median solving time of the experiments was 11 seconds). Tests on public data were run on an internal cluster with sufficient memory (up to 128GB for model training, CE generation requires notably less) and 16 CPUs, AMD EPYC 7543 or Intel Xeon Scalable Gold 6146, based on their availability.
% \jirimargincomment{runtime specs}
\subsection{Results}
% \jiricomment{mean is more difficult than sum}
We now compare the different variants of our modified LiCE for recommenders. We focus on the ability to provide globally optimal solutions and on the trade-off between plausibility and similarity.
Firstly, considering that MIO has a reputation for being slow to solve, we wish to establish how often our method obtains
% the global optimum
a solution
of the MIO formulation. In Table \ref{tab:rates}, we show the proportion of success in CE generation. Interestingly, this depends on multiple criteria. Firstly, note that without an SPN, we obtain a CE for each factual, irrespective of the input format or validity constraint ($k$-th vs score). Secondly, note that the thresholding variant fails often, except for the disjunction network. In particular, 98.8\% of the failures of the Threshold variant are cases where the model has been shown to be infeasible, likely due to the likelihood threshold being too prohibitive. %This is caused partly by the SPN format and partly by the threshold being too restrictive (nearly all failed generations are due to the formulation being infeasible), given the granularity of the trained SPN.
Thirdly, notice that using only score thresholding leads to a higher success rate, although it does not always lead to the explained item ranking beyond the $k$-th position. Indeed, approximately 30\% of CEs generated with the score threshold remain among top-$k$ recommendations.
Interestingly, this proportion is almost the same for $k$ equal to 5 and 10.
Finally, we see that the decisions of a recommender trained on ratings (rather than binarizations) seem to be more challenging to explain. %on the optimizing variant.
We can conclude that the disjunction aggregator function is the easiest to solve, and the learned SPN seems to be the least restrictive.
Thus, we restrict our focus to the disjunction aggregator on the binarized input vectors. In Table \ref{tab:stats}, we present the results on Plausibility and Similarity, measured by the log-likelihood and $\ell_1$ distance. Overall, the results correspond to the results of the original LiCE, where we see an improvement in plausibility, offset by some decrease in similarity.
% Notice especially the Dateio results with categories, where each method finds counterfactuals consisting of changing a single input, while achieving improved likelihood on average.
% \jirimargincomment{is this the own item id? It is probably caused by different number of the values over which we do the mean..}
\section{A User Study}
% We have also evaluated our method on the private Dateio dataset, using the production recommender system.
% % which would be meaningless to compare performance-wise.
% Its performance roughly corresponds to the performance on the public datasets. More importantly, we used the method to perform a study, where 20 customers were shown the explanations for up to 20 recommended merchants, constrained to a 5-position decrease in ranking.
% Each customer was presented with 5 different counterfactuals, each generated by a different variant. The customer was then asked to evaluate the subjective quality of the explanation. In the end, each subject was asked whether this improved their understanding of the model.
% CEs of all tested variants were considered understandable, and they helped strengthen users' trust in the system's reliability. The two variants considered to return the best CEs were a variant with a score decrease and no SPN, and a variant with a guaranteed 5-position drop with a disjunction SPN on samples aggregated by tags. The former was preferred because of the simplicity of the explanations, and the latter because of the ability to uncover more unexpected explanations.
% \jiricomment{describe the user-study results briefly}
% \subsection{Survey}
In addition to the numerical results above, we've performed a user study on merchant recommendation using private data from Dateio. There, the task is to suggest cash-back offers to customers, ranking the most relevant merchant offers first.
This relevance is estimated based on the user's transaction history, i.e., looking at which merchants were visited (and how often) by each customer.
The qualitative evaluation of the provided CEs is non-trivial. Transactions with suggested merchant and CE merchants may be related to various phenomena, such as the fied of business of the merchants, spatial patterns, or other user-specific shopping habits and preferences. Therefore, the best way to assess the quality of explanations is to collect feedback directly from the users, who can properly assess the relevance of counterfactual explanations. Simultaneously, we evaluated the quality of the recommender system, since it would not be efficient to discuss suggested merchants that are completely irrelevant to the user.
\begin{table}
\centering
\caption{Example of merchant suggestions using EASE.}
\begin{tabular}{clc}
\toprule
Rank & Merchant & EASE score \\
\midrule
1 & BILLA & 1.801893 \\
2 & McDonald's & 1.749221 \\
3 & Lidl & 1.700293 \\
4 & Albert & 1.641677 \\
\bottomrule
\end{tabular}
\label{tab:ease}
\end{table}
In order to test both the suggestions and their explanations, we created a survey that incorporates a personalized dashboard with transaction data. Merchant suggestions and CEs were generated based on the transaction history of each subject. Example ranking and counterfactual explanation can be seen in Table \ref{tab:ease} and Figure \ref{fig:dashboard}, respectively.
\begin{figure}
\centering
\fbox{\includegraphics[width=0.25\textwidth]{dashboard_1.png}} % Adjust width as needed
\caption{Example of a counterfactual explanation}
\label{fig:dashboard}
\end{figure}
First, the respondent provides a general assessment of merchant suggestions, primarily considering the relevance of included merchants and their relative order. Second, the CEs of suggested merchants are evaluated. We generate multiple alternative explanations for each merchant suggestion, corresponding to different validity constraints and SPNs used. All setups are described in Table \ref{tab:ce_approaches}, where the ``score decrease'' (in C, D, and E) means setting a threshold on the score that would correspond to a drop of 5 positions in the original ranking. The 5 position drop is guaranteed in the setups A and B. Category SPN uses merchant categories (see third column in Figure \ref{fig:dashboard}) and a sum as aggregator function. Binary tag SPN is an SPN using tags (more nuanced categories, with more than one associated value allowed) and disjunction as an aggregator function. None stands for not using an SPN at all. We impose an additional restriction that the originally suggested merchant cannot appear among the counterfactual merchant(s). This forces LiCE to find a more complex explanation\footnote{Otherwise, the explanation could say that merchant $i$ was recommended since the customer has transactions with merchant $i$, which is not a desired outcome here.}.
Overall, the goal is to assess whether the LiCE can be applied in merchant recommendatoin and to choose the best configuration to find the explanations.
\begin{table}[t]
\centering
\caption{Setups used int the user study.}
\begin{tabular}{clll}
\toprule
& Method & Condition & SPN variant used \\
\midrule
A & below n-th position & 5 positions drop & binary tag SPN \\
B & below n-th position & 5 positions dorp & none \\
C & below a given score & score decrease & none \\
% C & below a given score & score decrease\protect\footnotemark[2] & none \\
D & below a given score & score decrease & category SPN \\
E & below a given score & score decrease & binary tag SPN \\
\bottomrule
\end{tabular}
\label{tab:ce_approaches}
\end{table}
% \footnotetext[2]{Setting a score threshold that corresponds to a 5 position decrease.}
In total, 20 respondents with various demographic and socio-economic backgrounds took part in the survey, each of them evaluating approximately 10-20 merchant suggestions and corresponding counterfactual explanations. As seen in Table \ref{tab:resp_stats}, the median respondent has approximately 460 transactions within the last year, distributed among 113 merchants and 19 categories, which suggests a frequent and versatile usage of payment card.
\begin{table}[h]
\centering
\caption{Transaction history of respondents (one year period)}
\begin{tabular}{lrrrr}
\toprule
{} & transactions & merchants & categories & tags \\
\midrule
min & 41.0 & 19.0 & 7.0 & 22.0 \\
median & 465.0 & 113.0 & 19.0 & 92.0 \\
max & 1903.0 & 290.0 & 22.0 & 148.0 \\
\bottomrule
\end{tabular}
\label{tab:resp_stats}
\end{table}
\subsection{Results}
Table \ref{tab:comp_stats} presents the statistics of various CE approaches. One can observe significant differences in the run-time of finding a CE. The approach C performs the fastest and provides a solution for most cases. Moreover, it produces explanations using the fewest merchants, on average. That could contribute to more concise and clear explanations. Its lack of use of an SPN would however suggest possibly lower plausibility.
\begin{table}[h]
\centering
\caption{Computation statistics}
\begin{tabular}{lrrr}
\toprule
& Mean duration & Proportion obtained & Mean CE size\\
\midrule
A & 35.261 s & 0.786 & 4.415 merchants \\
B & 24.938 s & 0.774 & 4.037 merchants \\
C & \textbf{0.093 s} & \textbf{0.905} & \textbf{3.045} merchants \\
D & 13.647 s & 0.900 & 3.410 merchants \\
E & 0.312 s & \textbf{0.905} & 3.503 merchants \\
\bottomrule
\end{tabular}
\label{tab:comp_stats}
\end{table}
\subsubsection{Survey results}
The evaluation of merchant suggestions\footnote{Q: Evaluate the list of recommended merchants -- How well does it correspond to your shopping preferences? (relevance of suggested merchants, their order, ..)} shows that the average respondent rates them 7.4 points out of 10. All respondents are in general satisfied and provide a positive feedback, however, some of them point out that certain merchants, especially the biggest grocery store producers, rank undesirably high.
This might be related to the fact that the current recommender does not reflect the frequency of purchases. Therefore, a global supermarket chain, which appears in the transaction history of many users and is likely to obtain a high score, might rank among the top positions even if the user only makes a single transaction with this merchant. This insight was important for further development of the recommenders.
\begin{figure}[h] % 'h' means here
\centering
\includegraphics[width=0.5\textwidth]{points.png} % Adjust width as needed
\caption{Evaluation of CE explanations in terms of how sensible are they to users. Scale is from makes no sense (1) to makes prefect sense (6).}
\label{fig:points}
\end{figure}
Considering the evaluation of CEs, the results can be found in Figure \ref{fig:points}. In general, respondents rate the explanations rather positively, as the average score\footnote{The interpretation of awarded points: 1 - 2 = does not make sense, I do not agree; 3 - 4 = ok, but I also see merchants that might not be there; 5 - 6 = makes a perfect sense.} of every approach is more than 4 points out of 6, suggesting only a few confusing explanations. Since the Friedman test shows that there exist statistically significant differences between particular approaches\footnote{$H_0$: No stat. difference between approaches; p-value = $0.0088$}, multiple Wilcoxon signed-rank tests were performed and show that approach C is significantly better than B, D, and E, based on their pairwise comparison\footnote{$H_0$: the approaches do not statistically differ; the list of adjusted p-values (applying Bonferroni correction): A vs B: 1.0, A vs C: 0.472, A vs D: 1.0, A vs E: 1.0, B vs C: 0.022, B vs D: 1.0, B vs E: 1.0, C vs D: 0.009, C vs E: 0.006, D vs E: 1.0}. From a statistical perspective, the approach A, which yields a slightly lower average score than C, can be considered equivalent.
As in the case of merchant suggestion evaluation, respondents mention a possible bias towards big merchants, which frequently appear among the CE merchants, despite having a low number of transactions. This might be resolved by redesigning the recommender system to account for this. On the other hand, respondents emphasize that when it comes to more niche merchants, the explanations work significantly better and provide more interesting information. They also appreciate the geographical patterns that the CE can provide, interconnecting seemingly unrelated merchants whose branches are, however, situated close to each other.
Based on the feedback provided, respondents prefer CEs with a lower number of merchants, since they consider them more clear and concise. However, in some cases, the CE provides a complex list of many merchants that characterize the shopping habits of the respondent. Such CE might receive a high score even though it entails multiple merchants.
Finally, the respondents evaluate the overall contribution of LiCE. When deciding whether they would like to see it in a banking app\footnote{Q: Would you like to have the explanation algorithm in your mobile banking app (e.g., when displaying merchants with cash-backs)?}, the responses either show unequivocal support or disapproval, resulting in the mean score of 7 points out of 10. The final question of the survey is based on the evaluation of statement presented in Figure \ref{fig:contribution}, using the Likert scale with 7 response values ranging from Strongly disagree to Strongly agree. As the figure shows, the majority of the respondents agree that the CE framework helps to make recommender decisions more transparent and understandable, whereas \textit{no one disagrees} with the statement.
\begin{figure} % 'h' means here
\centering
\includegraphics[width=0.4\textwidth]{ce_contribution.png} % Adjust width as needed
\caption{Q: After seeing the explanations, I better understand the merchant-recommendation process, and I consider it more reliable.}
\label{fig:contribution}
\end{figure}
% \begin{figure} % 'h' means here
% \centering
% \includegraphics[width=0.5\textwidth]{ce_contribution_alt.png} % Adjust width as needed
% \caption{Contribution and usefulness of CE explanations}
% \label{fig:contribution2}
% \end{figure}
% From the Dateio perspective, the CE results are used internally to improve the EASE recommender system and to eliminate the undesired trends reported in the feedback of the respondents. The overall results confirm that EASE is an appropriate tool for this use case and provides notable help for the further development of our product.
\section{Conclusion}
% Based on the survey, it is possible to conclude that EASE is a useful method to generate personalized merchant suggestions for users of a banking app. Moreover, LiCE is an efficient tool for explaining merchant suggestions and increasing users' trust and positive impressions. When comparing several approaches for generating CE, the simplest and most efficient method provides the best evaluations of respondents, primarily because of the simplicity of the generated explanations.
% \section{Conclusion}
We have considered the likelihood of counterfactual explanations of a recommender system and extended the LiCE method to recommender systems. We tested multiple variants and found that, especially on binary attributes, our method can reliably return explanations faithful to the underlying model. We have also supported the claims of usefulness of the method by a user study.
%
Assuming that the explanations were provided only upon request, even the run-time of open-source solvers may be bearable.
Finally, obtaining plausible counterfactuals could also be useful for an in-distribution data-augmentation \citep{wang_counterfactual_2021} Counterfactual Collaborative Reasoning \citep{ji_counterfactual_2023} or in studying bias in RS \citep{bhadani_biases_2021,felicioni_enhancing_2022,wang_counterfactual_2024}.
% Finally, the proposed approach enables the model owner to specify various other behavior, such as explaining ranking restricted to some set of items, e.g., single category.
% An example of a floating figure using the graphicx package.
% Note that \label must occur AFTER (or within) \caption.
% For figures, \caption should occur after the \includegraphics.
% Note that IEEEtran v1.7 and later has special internal code that
% is designed to preserve the operation of \label within \caption
% even when the captionsoff option is in effect. However, because
% of issues like this, it may be the safest practice to put all your
% \label just after \caption rather than within \caption{}.
%
% Reminder: the "draftcls" or "draftclsnofoot", not "draft", class
% option should be used if it is desired that the figures are to be
% displayed while in draft mode.
%
%\begin{figure}[!t]
%\centering
%\includegraphics[width=2.5in]{myfigure}
% where an .eps filename suffix will be assumed under latex,
% and a .pdf suffix will be assumed for pdflatex; or what has been declared
% via \DeclareGraphicsExtensions.
%\caption{Simulation results for the network.}
%\label{fig_sim}
%\end{figure}
% Note that the IEEE typically puts floats only at the top, even when this
% results in a large percentage of a column being occupied by floats.
% An example of a double column floating figure using two subfigures.
% (The subfig.sty package must be loaded for this to work.)
% The subfigure \label commands are set within each subfloat command,
% and the \label for the overall figure must come after \caption.
% \hfil is used as a separator to get equal spacing.
% Watch out that the combined width of all the subfigures on a
% line do not exceed the text width or a line break will occur.
%
%\begin{figure*}[!t]
%\centering
%\subfloat[Case I]{\includegraphics[width=2.5in]{box}%
%\label{fig_first_case}}
%\hfil
%\subfloat[Case II]{\includegraphics[width=2.5in]{box}%
%\label{fig_second_case}}
%\caption{Simulation results for the network.}
%\label{fig_sim}
%\end{figure*}
%
% Note that often IEEE papers with subfigures do not employ subfigure
% captions (using the optional argument to \subfloat[]), but instead will
% reference/describe all of them (a), (b), etc., within the main caption.
% Be aware that for subfig.sty to generate the (a), (b), etc., subfigure
% labels, the optional argument to \subfloat must be present. If a
% subcaption is not desired, just leave its contents blank,
% e.g., \subfloat[].
% An example of a floating table. Note that, for IEEE style tables, the
% \caption command should come BEFORE the table and, given that table
% captions serve much like titles, are usually capitalized except for words
% such as a, an, and, as, at, but, by, for, in, nor, of, on, or, the, to
% and up, which are usually not capitalized unless they are the first or
% last word of the caption. Table text will default to \footnotesize as
% the IEEE normally uses this smaller font for tables.
% The \label must come after \caption as always.
%
%\begin{table}[!t]
%% increase table row spacing, adjust to taste
%\renewcommand{\arraystretch}{1.3}
% if using array.sty, it might be a good idea to tweak the value of
% \extrarowheight as needed to properly center the text within the cells
%\caption{An Example of a Table}
%\label{table_example}
%\centering
%% Some packages, such as MDW tools, offer better commands for making tables
%% than the plain LaTeX2e tabular which is used here.
%\begin{tabular}{|c||c|}
%\hline
%One & Two\\
%\hline
%Three & Four\\
%\hline
%\end{tabular}
%\end{table}
% Note that the IEEE does not put floats in the very first column
% - or typically anywhere on the first page for that matter. Also,
% in-text middle ("here") positioning is typically not used, but it
% is allowed and encouraged for Computer Society conferences (but
% not Computer Society journals). Most IEEE journals/conferences use
% top floats exclusively.
% Note that, LaTeX2e, unlike IEEE journals/conferences, places
% footnotes above bottom floats. This can be corrected via the
% \fnbelowfloat command of the stfloats package.
% \section{Conclusion}
% The conclusion goes here.
% conference papers do not normally have an appendix
% use section* for acknowledgment
\ifCLASSOPTIONcompsoc
% The Computer Society usually uses the plural form
\section*{Acknowledgments}
\else
% regular IEEE prefers the singular form
\section*{Acknowledgment}
\fi
This work has received funding from the European Union’s Horizon Europe research and innovation programme under grant agreement No. 101070568.
% trigger a \newpage just before the given reference
% number - used to balance the columns on the last page
% adjust value as needed - may need to be readjusted if
% the document is modified later
%\IEEEtriggeratref{8}
% The "triggered" command can be changed if desired:
%\IEEEtriggercmd{\enlargethispage{-5in}}
% references section
% can use a bibliography generated by BibTeX as a .bbl file
% BibTeX documentation can be easily obtained at:
% http://mirror.ctan.org/biblio/bibtex/contrib/doc/
% The IEEEtran BibTeX style support page is at:
% http://www.michaelshell.org/tex/ieeetran/bibtex/
\bibliographystyle{IEEEtran}
% argument is your BibTeX string definitions and bibliography database(s)
\bibliography{references}
% that's all folks
\end{document}
% Generated by IEEEtran.bst, version: 1.14 (2015/08/26)
\begin{thebibliography}{10}
\providecommand{\url}[1]{#1}
\csname url@samestyle\endcsname
\providecommand{\newblock}{\relax}
\providecommand{\bibinfo}[2]{#2}
\providecommand{\BIBentrySTDinterwordspacing}{\spaceskip=0pt\relax}
\providecommand{\BIBentryALTinterwordstretchfactor}{4}
\providecommand{\BIBentryALTinterwordspacing}{\spaceskip=\fontdimen2\font plus
\BIBentryALTinterwordstretchfactor\fontdimen3\font minus \fontdimen4\font\relax}
\providecommand{\BIBforeignlanguage}[2]{{%
\expandafter\ifx\csname l@#1\endcsname\relax
\typeout{** WARNING: IEEEtran.bst: No hyphenation pattern has been}%
\typeout{** loaded for the language `#1'. Using the pattern for}%
\typeout{** the default language instead.}%
\else
\language=\csname l@#1\endcsname
\fi
#2}}
\providecommand{\BIBdecl}{\relax}
\BIBdecl
\bibitem{herlocker_explaining_2000}
\BIBentryALTinterwordspacing
J.~L. Herlocker, J.~A. Konstan, and J.~Riedl, ``\BIBforeignlanguage{en}{Explaining collaborative filtering recommendations},'' in \emph{\BIBforeignlanguage{en}{Proceedings of the 2000 {ACM} conference on {Computer} supported cooperative work}}.\hskip 1em plus 0.5em minus 0.4em\relax Philadelphia Pennsylvania USA: ACM, Dec. 2000, pp. 241--250. [Online]. Available: \url{https://dl.acm.org/doi/10.1145/358916.358995}
\BIBentrySTDinterwordspacing
\bibitem{shang_why_2022}
\BIBentryALTinterwordspacing
R.~Shang, K.~J.~K. Feng, and C.~Shah, ``\BIBforeignlanguage{en}{Why {Am} {I} {Not} {Seeing} {It}? {Understanding} {Users}’ {Needs} for {Counterfactual} {Explanations} in {Everyday} {Recommendations}},'' in \emph{\BIBforeignlanguage{en}{2022 {ACM} {Conference} on {Fairness}, {Accountability}, and {Transparency}}}.\hskip 1em plus 0.5em minus 0.4em\relax Seoul Republic of Korea: ACM, Jun. 2022, pp. 1330--1340. [Online]. Available: \url{https://dl.acm.org/doi/10.1145/3531146.3533189}
\BIBentrySTDinterwordspacing
\bibitem{nunes_systematic_2017}
\BIBentryALTinterwordspacing
I.~Nunes and D.~Jannach, ``\BIBforeignlanguage{en}{A systematic review and taxonomy of explanations in decision support and recommender systems},'' \emph{\BIBforeignlanguage{en}{User Modeling and User-Adapted Interaction}}, vol.~27, no.~3, pp. 393--444, Dec. 2017. [Online]. Available: \url{https://doi.org/10.1007/s11257-017-9195-0}
\BIBentrySTDinterwordspacing
\bibitem{mcinerney_explore_2018}
\BIBentryALTinterwordspacing
J.~McInerney, B.~Lacker, S.~Hansen, K.~Higley, H.~Bouchard, A.~Gruson, and R.~Mehrotra, ``Explore, exploit, and explain: personalizing explainable recommendations with bandits,'' in \emph{Proceedings of the 12th {ACM} {Conference} on {Recommender} {Systems}}, ser. {RecSys} '18.\hskip 1em plus 0.5em minus 0.4em\relax New York, NY, USA: Association for Computing Machinery, Sep. 2018, pp. 31--39. [Online]. Available: \url{https://doi.org/10.1145/3240323.3240354}
\BIBentrySTDinterwordspacing
\bibitem{gdpr}
{Regulation (EU) 2016/679}, ``Regulation ({EU}) 2016/679 of the {European} {Parliament} and of the {Council}. of 27 {April} 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such data, and repealing {Directive} 95/46/{EC} ({General} {Data} {Protection} {Regulation}),'' 2016, place: OJ L 119, 4.5.2016, p. 1–88.
\bibitem{CCPA}
\BIBentryALTinterwordspacing
``California consumer privacy act (ccpa),'' \emph{Retrieved from State of California Department of Justice}, 2022. [Online]. Available: \url{https://oag.ca.gov/privacy/ccpa}
\BIBentrySTDinterwordspacing
\bibitem{ai-act}
{Regulation (EU) 2024/1689}, ``Regulation ({EU}) 2024/1689 of the {European} {Parliament} and of the {Council}. of 13 {June} 2024 laying down harmonised rules on artificial intelligence and amending {Regulations} ({EC}) {No} 300/2008, ({EU}) {No} 167/2013, ({EU}) {No} 168/2013, ({EU}) 2018/858, ({EU}) 2018/1139 and ({EU}) 2019/2144 and {Directives} 2014/90/{EU}, ({EU}) 2016/797 and ({EU}) 2020/1828 ({Artificial} {Intelligence} {Act}),'' 2024, place: OJ L, 2024/1689, 12.7.2024.
\bibitem{ge_survey_2024}
\BIBentryALTinterwordspacing
Y.~Ge, S.~Liu, Z.~Fu, J.~Tan, Z.~Li, S.~Xu, Y.~Li, Y.~Xian, and Y.~Zhang, ``\BIBforeignlanguage{en}{A {Survey} on {Trustworthy} {Recommender} {Systems}},'' \emph{\BIBforeignlanguage{en}{ACM Transactions on Recommender Systems}}, vol.~3, no.~2, pp. 1--68, Jun. 2024. [Online]. Available: \url{https://dl.acm.org/doi/10.1145/3652891}
\BIBentrySTDinterwordspacing
\bibitem{wachter_counterfactual_2017}
\BIBentryALTinterwordspacing
S.~Wachter, B.~Mittelstadt, and C.~Russell, ``\BIBforeignlanguage{en}{Counterfactual {Explanations} {Without} {Opening} the {Black} {Box}: {Automated} {Decisions} and the {GDPR}},'' \emph{\BIBforeignlanguage{en}{SSRN Electronic Journal}}, 2017. [Online]. Available: \url{https://www.ssrn.com/abstract=3063289}
\BIBentrySTDinterwordspacing
\bibitem{gdpr_supports_ce_case}
{EU Court of Justice}, ``{{CK}} v {{Dun}} \& {{Bradstreet Austria GmbH}} and {{Magistrat}} der {{Stadt Wien}},'' Feb. 2025.
\bibitem{guidotti_counterfactual_2022}
\BIBentryALTinterwordspacing
R.~Guidotti, ``\BIBforeignlanguage{en}{Counterfactual explanations and how to find them: literature review and benchmarking},'' \emph{\BIBforeignlanguage{en}{Data Mining and Knowledge Discovery}}, Apr. 2022. [Online]. Available: \url{https://doi.org/10.1007/s10618-022-00831-6}
\BIBentrySTDinterwordspacing
\bibitem{baklanov_ceers_2024}
\BIBentryALTinterwordspacing
M.~Baklanov, ``{CEERS}: {Counterfactual} {Evaluations} of {Explanations} in {Recommender} {Systems},'' in \emph{Proceedings of the 18th {ACM} {Conference} on {Recommender} {Systems}}, ser. {RecSys} '24.\hskip 1em plus 0.5em minus 0.4em\relax New York, NY, USA: Association for Computing Machinery, Oct. 2024, pp. 1323--1329. [Online]. Available: \url{https://dl.acm.org/doi/10.1145/3640457.3688015}
\BIBentrySTDinterwordspacing
\bibitem{nemecek2025generating}
\BIBentryALTinterwordspacing
J.~Němeček, T.~Pevný, and J.~Mareček, ``Generating likely counterfactuals using sum-product networks,'' in \emph{The thirteenth international conference on learning representations}, 2025. [Online]. Available: \url{https://openreview.net/forum?id=rGyi8NNqB0}
\BIBentrySTDinterwordspacing
\bibitem{wolsey_integer_2021}
L.~A. Wolsey, \emph{Integer programming}, second edition~ed.\hskip 1em plus 0.5em minus 0.4em\relax Hoboken, NJ: Wiley, 2021.
\bibitem{kochProgressMathematicalProgramming2022}
T.~Koch, T.~Berthold, J.~Pedersen, and C.~Vanaret, ``Progress in mathematical programming solvers from 2001 to 2020,'' \emph{EURO Journal on Computational Optimization}, vol.~10, p. 100031, Jan. 2022.
\bibitem{poon_sum-product_2011}
\BIBentryALTinterwordspacing
H.~Poon and P.~Domingos, ``Sum-product networks: {A} new deep architecture,'' in \emph{2011 {IEEE} {International} {Conference} on {Computer} {Vision} {Workshops} ({ICCV} {Workshops})}, Nov. 2011, pp. 689--690. [Online]. Available: \url{https://ieeexplore.ieee.org/abstract/document/6130310}
\BIBentrySTDinterwordspacing
\bibitem{NEURIPS2023_6b61c278}
\BIBentryALTinterwordspacing
Z.~Yu, M.~Trapp, and K.~Kersting, ``Characteristic circuits,'' in \emph{Advances in Neural Information Processing Systems}, A.~Oh, T.~Naumann, A.~Globerson, K.~Saenko, M.~Hardt, and S.~Levine, Eds., vol.~36.\hskip 1em plus 0.5em minus 0.4em\relax Curran Associates, Inc., 2023, pp. 34\,074--34\,086. [Online]. Available: \url{https://proceedings.neurips.cc/paper_files/paper/2023/file/6b61c278e483954fee502b49fe71cd14-Paper-Conference.pdf}
\BIBentrySTDinterwordspacing
\bibitem{sidheekh2023probabilistic}
S.~Sidheekh, K.~Kersting, and S.~Natarajan, ``Probabilistic flow circuits: Towards unified deep models for tractable probabilistic inference,'' in \emph{Uncertainty in Artificial Intelligence}.\hskip 1em plus 0.5em minus 0.4em\relax PMLR, 2023, pp. 1964--1973.
\bibitem{mcsherry_explanation_2005}
\BIBentryALTinterwordspacing
D.~Mcsherry, ``\BIBforeignlanguage{en}{Explanation in {Recommender} {Systems}},'' \emph{\BIBforeignlanguage{en}{Artificial Intelligence Review}}, vol.~24, no.~2, pp. 179--197, Oct. 2005. [Online]. Available: \url{https://doi.org/10.1007/s10462-005-4612-x}
\BIBentrySTDinterwordspacing
\bibitem{friedrich_taxonomy_2011}
\BIBentryALTinterwordspacing
G.~Friedrich and M.~Zanker, ``\BIBforeignlanguage{en}{A {Taxonomy} for {Generating} {Explanations} in {Recommender} {Systems}},'' \emph{\BIBforeignlanguage{en}{AI Magazine}}, vol.~32, no.~3, pp. 90--98, 2011, \_eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1609/aimag.v32i3.2365. [Online]. Available: \url{https://onlinelibrary.wiley.com/doi/abs/10.1609/aimag.v32i3.2365}
\BIBentrySTDinterwordspacing
\bibitem{vultureanu-albisi_survey_2022}
\BIBentryALTinterwordspacing
A.~Vultureanu-Albişi and C.~Bădică, ``\BIBforeignlanguage{en}{A survey on effects of adding explanations to recommender systems},'' \emph{\BIBforeignlanguage{en}{Concurrency and Computation: Practice and Experience}}, vol.~34, no.~20, p. e6834, 2022, \_eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1002/cpe.6834. [Online]. Available: \url{https://onlinelibrary.wiley.com/doi/abs/10.1002/cpe.6834}
\BIBentrySTDinterwordspacing
\bibitem{wardatzky_evaluating_2024}
\BIBentryALTinterwordspacing
K.~Wardatzky, ``Evaluating the {Pros} and {Cons} of {Recommender} {Systems} {Explanations},'' in \emph{Proceedings of the 18th {ACM} {Conference} on {Recommender} {Systems}}, ser. {RecSys} '24.\hskip 1em plus 0.5em minus 0.4em\relax New York, NY, USA: Association for Computing Machinery, Oct. 2024, pp. 1302--1307. [Online]. Available: \url{https://dl.acm.org/doi/10.1145/3640457.3688011}
\BIBentrySTDinterwordspacing
\bibitem{ghazimatin_prince_2020}
\BIBentryALTinterwordspacing
A.~Ghazimatin, O.~Balalau, R.~Saha~Roy, and G.~Weikum, ``\BIBforeignlanguage{en}{{PRINCE}: {Provider}-side {Interpretability} with {Counterfactual} {Explanations} in {Recommender} {Systems}},'' in \emph{\BIBforeignlanguage{en}{Proceedings of the 13th {International} {Conference} on {Web} {Search} and {Data} {Mining}}}.\hskip 1em plus 0.5em minus 0.4em\relax Houston TX USA: ACM, Jan. 2020, pp. 196--204. [Online]. Available: \url{https://dl.acm.org/doi/10.1145/3336191.3371824}
\BIBentrySTDinterwordspacing
\bibitem{tran_counterfactual_2021}
\BIBentryALTinterwordspacing
K.~H. Tran, A.~Ghazimatin, and R.~Saha~Roy, ``\BIBforeignlanguage{en}{Counterfactual {Explanations} for {Neural} {Recommenders}},'' in \emph{\BIBforeignlanguage{en}{Proceedings of the 44th {International} {ACM} {SIGIR} {Conference} on {Research} and {Development} in {Information} {Retrieval}}}.\hskip 1em plus 0.5em minus 0.4em\relax Virtual Event Canada: ACM, Jul. 2021, pp. 1627--1631. [Online]. Available: \url{https://dl.acm.org/doi/10.1145/3404835.3463005}
\BIBentrySTDinterwordspacing
\bibitem{xu_learning_2021}
S.~Xu, Y.~Li, S.~Liu, Z.~Fu, Y.~Ge, X.~Chen, and Y.~Zhang, ``\BIBforeignlanguage{en}{Learning {Causal} {Explanations} for {Recommendation}},'' 2021.
\bibitem{he_prototype-guided_2023}
M.~He, J.~Wang, B.~An, and H.~Wen, ``\BIBforeignlanguage{en}{Prototype-{Guided} {Counterfactual} {Explanations} via {Variational} {Auto}-encoder for {Recommendation}},'' in \emph{\BIBforeignlanguage{en}{Machine {Learning} and {Knowledge} {Discovery} in {Databases}: {Applied} {Data} {Science} and {Demo} {Track}}}, G.~De~Francisci~Morales, C.~Perlich, N.~Ruchansky, N.~Kourtellis, E.~Baralis, and F.~Bonchi, Eds.\hskip 1em plus 0.5em minus 0.4em\relax Cham: Springer Nature Switzerland, 2023, pp. 652--668.
\bibitem{tan_counterfactual_2021}
\BIBentryALTinterwordspacing
J.~Tan, S.~Xu, Y.~Ge, Y.~Li, X.~Chen, and Y.~Zhang, ``\BIBforeignlanguage{en}{Counterfactual {Explainable} {Recommendation}},'' in \emph{\BIBforeignlanguage{en}{Proceedings of the 30th {ACM} {International} {Conference} on {Information} \& {Knowledge} {Management}}}.\hskip 1em plus 0.5em minus 0.4em\relax Virtual Event Queensland Australia: ACM, Oct. 2021, pp. 1784--1793. [Online]. Available: \url{https://dl.acm.org/doi/10.1145/3459637.3482420}
\BIBentrySTDinterwordspacing
\bibitem{Tan_2022}
\BIBentryALTinterwordspacing
J.~Tan, S.~Geng, Z.~Fu, Y.~Ge, S.~Xu, Y.~Li, and Y.~Zhang, ``Learning and evaluating graph neural network explanations based on counterfactual and factual reasoning,'' in \emph{Proceedings of the {ACM} web conference 2022}, ser. Www ’22.\hskip 1em plus 0.5em minus 0.4em\relax ACM, Apr. 2022, pp. 1018--1027. [Online]. Available: \url{http://dx.doi.org/10.1145/3485447.3511948}
\BIBentrySTDinterwordspacing
\bibitem{chen_grease_2022}
\BIBentryALTinterwordspacing
Z.~Chen, F.~Silvestri, J.~Wang, Y.~Zhang, Z.~Huang, H.~Ahn, and G.~Tolomei, ``\BIBforeignlanguage{en}{{GREASE}: {Generate} {Factual} and {Counterfactual} {Explanations} for {GNN}-based {Recommendations}},'' Aug. 2022, arXiv:2208.04222 [cs]. [Online]. Available: \url{http://arxiv.org/abs/2208.04222}
\BIBentrySTDinterwordspacing
\bibitem{kaffes_model-agnostic_2021}
\BIBentryALTinterwordspacing
V.~Kaffes, D.~Sacharidis, and G.~Giannopoulos, ``\BIBforeignlanguage{en}{Model-{Agnostic} {Counterfactual} {Explanations} of {Recommendations}},'' in \emph{\BIBforeignlanguage{en}{Proceedings of the 29th {ACM} {Conference} on {User} {Modeling}, {Adaptation} and {Personalization}}}.\hskip 1em plus 0.5em minus 0.4em\relax Utrecht Netherlands: ACM, Jun. 2021, pp. 280--285. [Online]. Available: \url{https://dl.acm.org/doi/10.1145/3450613.3456846}
\BIBentrySTDinterwordspacing
\bibitem{zhong_shap-enhanced_2022}
\BIBentryALTinterwordspacing
J.~Zhong and E.~Negre, ``Shap-enhanced counterfactual explanations for recommendations,'' in \emph{Proceedings of the 37th {ACM}/{SIGAPP} {Symposium} on {Applied} {Computing}}, ser. {SAC} '22.\hskip 1em plus 0.5em minus 0.4em\relax New York, NY, USA: Association for Computing Machinery, May 2022, pp. 1365--1372. [Online]. Available: \url{https://dl.acm.org/doi/10.1145/3477314.3507029}
\BIBentrySTDinterwordspacing
\bibitem{ge2024survey}
Y.~Ge, S.~Liu, Z.~Fu, J.~Tan, Z.~Li, S.~Xu, Y.~Li, Y.~Xian, and Y.~Zhang, ``A survey on trustworthy recommender systems,'' \emph{ACM Transactions on Recommender Systems}, vol.~3, no.~2, pp. 1--68, 2024.
\bibitem{he_neural_2017}
\BIBentryALTinterwordspacing
X.~He, L.~Liao, H.~Zhang, L.~Nie, X.~Hu, and T.-S. Chua, ``Neural {Collaborative} {Filtering},'' in \emph{Proceedings of the 26th {International} {Conference} on {World} {Wide} {Web}}, ser. {WWW} '17.\hskip 1em plus 0.5em minus 0.4em\relax Republic and Canton of Geneva, CHE: International World Wide Web Conferences Steering Committee, Apr. 2017, pp. 173--182. [Online]. Available: \url{https://doi.org/10.1145/3038912.3052569}
\BIBentrySTDinterwordspacing
\bibitem{steck_embarrassingly_2019}
\BIBentryALTinterwordspacing
H.~Steck, ``\BIBforeignlanguage{en}{Embarrassingly {Shallow} {Autoencoders} for {Sparse} {Data}},'' in \emph{\BIBforeignlanguage{en}{The {World} {Wide} {Web} {Conference}}}.\hskip 1em plus 0.5em minus 0.4em\relax San Francisco CA USA: ACM, May 2019, pp. 3251--3257. [Online]. Available: \url{https://dl.acm.org/doi/10.1145/3308558.3313710}
\BIBentrySTDinterwordspacing
\bibitem{vancura_deep_2021}
V.~Vančura and P.~Kordík, ``\BIBforeignlanguage{en}{Deep {Variational} {Autoencoder} with {Shallow} {Parallel} {Path} for {Top}-{N} {Recommendation} ({VASP})},'' in \emph{\BIBforeignlanguage{en}{Artificial {Neural} {Networks} and {Machine} {Learning} – {ICANN} 2021}}, I.~Farkaš, P.~Masulli, S.~Otte, and S.~Wermter, Eds.\hskip 1em plus 0.5em minus 0.4em\relax Cham: Springer International Publishing, 2021, pp. 138--149.
\bibitem{liang_variational_2018}
\BIBentryALTinterwordspacing
D.~Liang, R.~G. Krishnan, M.~D. Hoffman, and T.~Jebara, ``Variational {Autoencoders} for {Collaborative} {Filtering},'' in \emph{Proceedings of the 2018 {World} {Wide} {Web} {Conference}}, ser. {WWW} '18.\hskip 1em plus 0.5em minus 0.4em\relax Republic and Canton of Geneva, CHE: International World Wide Web Conferences Steering Committee, Apr. 2018, pp. 689--698. [Online]. Available: \url{https://dl.acm.org/doi/10.1145/3178876.3186150}
\BIBentrySTDinterwordspacing
\bibitem{anderson_strong_2020}
\BIBentryALTinterwordspacing
R.~Anderson, J.~Huchette, W.~Ma, C.~Tjandraatmadja, and J.~P. Vielma, ``\BIBforeignlanguage{en}{Strong mixed-integer programming formulations for trained neural networks},'' \emph{\BIBforeignlanguage{en}{Mathematical Programming}}, vol. 183, no.~1, pp. 3--39, Sep. 2020. [Online]. Available: \url{https://doi.org/10.1007/s10107-020-01474-5}
\BIBentrySTDinterwordspacing
\bibitem{bynum_pyomo_2021}
M.~L. Bynum, G.~A. Hackebeil, W.~E. Hart, C.~D. Laird, B.~L. Nicholson, J.~D. Siirola, J.-P. Watson, D.~L. Woodruff, and {others}, \emph{Pyomo - {Optimization} {Modeling} in {Python}, 3rd {Edition}}.\hskip 1em plus 0.5em minus 0.4em\relax Springer, 2021, vol.~67.
\bibitem{ning_slim_2011}
\BIBentryALTinterwordspacing
X.~Ning and G.~Karypis, ``{SLIM}: {Sparse} {Linear} {Methods} for {Top}-{N} {Recommender} {Systems},'' in \emph{2011 {IEEE} 11th {International} {Conference} on {Data} {Mining}}, Dec. 2011, pp. 497--506, iSSN: 2374-8486. [Online]. Available: \url{https://ieeexplore.ieee.org/document/6137254}
\BIBentrySTDinterwordspacing
\bibitem{kim_enhancing_2019}
\BIBentryALTinterwordspacing
D.~Kim and B.~Suh, ``Enhancing {VAEs} for collaborative filtering: flexible priors \& gating mechanisms,'' in \emph{Proceedings of the 13th {ACM} {Conference} on {Recommender} {Systems}}, ser. {RecSys} '19.\hskip 1em plus 0.5em minus 0.4em\relax New York, NY, USA: Association for Computing Machinery, Sep. 2019, pp. 403--407. [Online]. Available: \url{https://dl.acm.org/doi/10.1145/3298689.3347015}
\BIBentrySTDinterwordspacing
\bibitem{yelp_data}
\BIBentryALTinterwordspacing
{Yelp}, ``\BIBforeignlanguage{en-US}{Yelp {Open} {Dataset}}.'' [Online]. Available: \url{https://business.yelp.com/data/resources/open-dataset/}
\BIBentrySTDinterwordspacing
\bibitem{amazon_review_data}
Y.~Hou, J.~Li, Z.~He, A.~Yan, X.~Chen, and J.~McAuley, ``Bridging language and items for retrieval and recommendation,'' \emph{arXiv preprint arXiv:2403.03952}, 2024.
\bibitem{netflix_data}
{James Bennett} and {Stan Lanning}, ``The {Netflix} {Prize},'' Aug. 2007.
\bibitem{mohammadi_are_2024}
\BIBentryALTinterwordspacing
A.~R. Mohammadi, A.~Peintner, M.~Müller, and E.~Zangerle, ``Are {We} {Explaining} the {Same} {Recommenders}? {Incorporating} {Recommender} {Performance} for {Evaluating} {Explainers},'' in \emph{Proceedings of the 18th {ACM} {Conference} on {Recommender} {Systems}}, ser. {RecSys} '24.\hskip 1em plus 0.5em minus 0.4em\relax New York, NY, USA: Association for Computing Machinery, Oct. 2024, pp. 1113--1118. [Online]. Available: \url{https://dl.acm.org/doi/10.1145/3640457.3691709}
\BIBentrySTDinterwordspacing
\bibitem{huangfuParallelizingDualRevised2018}
Q.~Huangfu and J.~A.~J. Hall, ``Parallelizing the dual revised simplex method,'' \emph{Mathematical Programming Computation}, vol.~10, no.~1, pp. 119--142, Mar. 2018.
\bibitem{wang_counterfactual_2021}
\BIBentryALTinterwordspacing
Z.~Wang, J.~Zhang, H.~Xu, X.~Chen, Y.~Zhang, W.~X. Zhao, and J.-R. Wen, ``Counterfactual {Data}-{Augmented} {Sequential} {Recommendation},'' in \emph{Proceedings of the 44th {International} {ACM} {SIGIR} {Conference} on {Research} and {Development} in {Information} {Retrieval}}, ser. {SIGIR} '21.\hskip 1em plus 0.5em minus 0.4em\relax New York, NY, USA: Association for Computing Machinery, Jul. 2021, pp. 347--356. [Online]. Available: \url{https://dl.acm.org/doi/10.1145/3404835.3462855}
\BIBentrySTDinterwordspacing
\bibitem{ji_counterfactual_2023}
\BIBentryALTinterwordspacing
J.~Ji, Z.~Li, S.~Xu, M.~Xiong, J.~Tan, Y.~Ge, H.~Wang, and Y.~Zhang, ``Counterfactual {Collaborative} {Reasoning},'' in \emph{Proceedings of the {Sixteenth} {ACM} {International} {Conference} on {Web} {Search} and {Data} {Mining}}, ser. {WSDM} '23.\hskip 1em plus 0.5em minus 0.4em\relax New York, NY, USA: Association for Computing Machinery, Feb. 2023, pp. 249--257. [Online]. Available: \url{https://dl.acm.org/doi/10.1145/3539597.3570464}
\BIBentrySTDinterwordspacing
\bibitem{bhadani_biases_2021}
\BIBentryALTinterwordspacing
S.~Bhadani, ``Biases in {Recommendation} {System},'' in \emph{Proceedings of the 15th {ACM} {Conference} on {Recommender} {Systems}}, ser. {RecSys} '21.\hskip 1em plus 0.5em minus 0.4em\relax New York, NY, USA: Association for Computing Machinery, Sep. 2021, pp. 855--859. [Online]. Available: \url{https://dl.acm.org/doi/10.1145/3460231.3473897}
\BIBentrySTDinterwordspacing
\bibitem{felicioni_enhancing_2022}
\BIBentryALTinterwordspacing
N.~Felicioni, ``Enhancing {Counterfactual} {Evaluation} and {Learning} for {Recommendation} {Systems},'' in \emph{Proceedings of the 16th {ACM} {Conference} on {Recommender} {Systems}}, ser. {RecSys} '22.\hskip 1em plus 0.5em minus 0.4em\relax New York, NY, USA: Association for Computing Machinery, Sep. 2022, pp. 739--741. [Online]. Available: \url{https://dl.acm.org/doi/10.1145/3523227.3547429}
\BIBentrySTDinterwordspacing
\bibitem{wang_counterfactual_2024}
\BIBentryALTinterwordspacing
X.~Wang, Q.~Li, D.~Yu, Q.~Li, and G.~Xu, ``Counterfactual {Explanation} for {Fairness} in {Recommendation},'' \emph{ACM Trans. Inf. Syst.}, vol.~42, no.~4, pp. 106:1--106:30, Mar. 2024. [Online]. Available: \url{https://dl.acm.org/doi/10.1145/3643670}
\BIBentrySTDinterwordspacing
\end{thebibliography}