ASAP: Unsupervised Post-training with Label Distribution Shift Adaptive Learning Rate

Extracted source text · arXiv 2508.13445

%%
%% This is file `sample-sigconf.tex',
%% generated with the docstrip utility.
%%
%% The original source files were:
%%
%% samples.dtx  (with options: `all,proceedings,bibtex,sigconf')
%% 
%% IMPORTANT NOTICE:
%% 
%% For the copyright see the source file.
%% 
%% Any modified versions of this file must be renamed
%% with new filenames distinct from sample-sigconf.tex.
%% 
%% For distribution of the original source see the terms
%% for copying and modification in the file samples.dtx.
%% 
%% This generated file may be distributed as long as the
%% original source files, as listed above, are part of the
%% same distribution. (The sources need not necessarily be
%% in the same archive or directory.)
%%
%%
%% Commands for TeXCount
%TC:macro \cite [option:text,text]
%TC:macro \citep [option:text,text]
%TC:macro \citet [option:text,text]
%TC:envir table 0 1
%TC:envir table* 0 1
%TC:envir tabular [ignore] word
%TC:envir displaymath 0 word
%TC:envir math 0 word
%TC:envir comment 0 0
%%
%% The first command in your LaTeX source must be the \documentclass
%% command.
%%
%% For submission and review of your manuscript please change the
%% command to \documentclass[manuscript, screen, review]{acmart}.
%%
%% When submitting camera ready or to TAPS, please change the command
%% to \documentclass[sigconf]{acmart} or whichever template is required
%% for your publication.
%%
%%

\documentclass[sigconf]{acmart}
\settopmatter{authorsperrow=4}
\usepackage[font=small,skip=2pt]{caption} % skip을 줄여 캡션과 그림 사이 간격 조절
\usepackage{subcaption}
\usepackage{pgffor}
\usepackage[toc,page,header]{appendix}
\usepackage{minitoc}
\usepackage{wrapfig}
\usepackage[utf8]{inputenc} % allow utf-8 input
\usepackage[T1]{fontenc}    % use 8-bit T1 fonts
\usepackage{hyperref}       % hyperlinks
\usepackage{url}            % simple URL typesetting
\usepackage{booktabs}       % professional-quality tables
\usepackage{amsfonts}       % blackboard math symbols
\usepackage{nicefrac}       % compact symbols for 1/2, etc.
\usepackage{microtype}      % microtypography
\usepackage{xcolor}         % colors
\usepackage{amsmath}
\usepackage{algorithmicx}
\usepackage{algorithm}
\usepackage{algpseudocode}
%\usepackage{amssymb}
\usepackage{booktabs}
\usepackage{multirow}
\usepackage{graphicx}
\usepackage{float}
\usepackage{enumitem}
\usepackage{tcolorbox}
\usepackage{graphicx}   % for \resizebox
\usepackage{colortbl}   % for \columncolor
%\documentclass[sigconf]{acmart}
%\documentclass[manuscript, screen, review]{acmart}
%%
%% \BibTeX command to typeset BibTeX logo in the docs
\AtBeginDocument{%
  \providecommand\BibTeX{{%
    Bib\TeX}}}

%% Rights management information.  This information is sent to you
%% when you complete the rights form.  These commands have SAMPLE
%% values in them; it is your responsibility as an author to replace
%% the commands and values with those provided to you when you
%% complete the rights form.
%\setcopyright{acmlicensed}
%\copyrightyear{2025}
%\acmYear{2025}
%\acmDOI{XXXXXXX.XXXXXXX}
%% These commands are for a PROCEEDINGS abstract or paper.
%\acmConference[CIKM '25]{The 34th ACM International Conference on Information and Knowledge Management}{November 10--14, 2025}{Seoul, Korea}
%\acmBooktitle{The 34th ACM International Conference on Information and Knowledge Management (CIKM '25), November 10--14,  2025, Seoul, Korea}
%%
%%  Uncomment \acmBooktitle if the title of the proceedings is different
%%  from ``Proceedings of ...''!
%%
%%\acmBooktitle{Woodstock '18: ACM Symposium on Neural Gaze Detection,
%%  June 03--05, 2018, Woodstock, NY}
%\acmISBN{978-1-4503-XXXX-X/2018/06}


%%
%% Submission ID.
%% Use this when submitting an article to a sponsored event. You'll
%% receive a unique submission ID from the organizers
%% of the event, and this ID should be used as the parameter to this command.
%%\acmSubmissionID{123-A56-BU3}

%%
%% For managing citations, it is recommended to use bibliography
%% files in BibTeX format.
%%
%% You can then either use BibTeX with the ACM-Reference-Format style,
%% or BibLaTeX with the acmnumeric or acmauthoryear sytles, that include
%% support for advanced citation of software artefact from the
%% biblatex-software package, also separately available on CTAN.
%%
%% Look at the sample-*-biblatex.tex files for templates showcasing
%% the biblatex styles.
%%

%%
%% The majority of ACM publications use numbered citations and
%% references.  The command \citestyle{authoryear} switches to the
%% "author year" style.
%%
%% If you are preparing content for an event
%% sponsored by ACM SIGGRAPH, you must use the "author year" style of
%% citations and references.
%% Uncommenting
%% the next command will enable that style.
%%\citestyle{acmauthoryear}

%%
%% end of the preamble, start of the body of the document source.

\copyrightyear{2025} 
\acmYear{2025} 
\setcopyright{cc}
\setcctype{by-nc-sa}
\acmConference[CIKM '25]{Proceedings of the 34th ACM International Conference on Information and Knowledge Management}{November 10--14, 2025}{Seoul, Republic of Korea}
\acmBooktitle{Proceedings of the 34th ACM International Conference on Information and Knowledge Management (CIKM '25),  November 10--14, 2025, Seoul, Republic of Korea}
\acmDOI{10.1145/3746252.3760830}
\acmISBN{979-8-4007-2040-6 /2025 /11}

\begin{document}

%%
%% The "title" command has an optional parameter,
%% allowing the author to define a "short title" to be used in page headers.
\title{ASAP: Unsupervised Post-training with Label Distribution Shift Adaptive Learning Rate
%ASAP: Post-training with Adaptive Learning Rate based on Label Shift Prediction
}

%%
%% The "author" command and its associated commands are used to define
%% the authors and their affiliations.
%% Of note is the shared affiliation of the first two authors, and the
%% "authornote" and "authornotemark" commands
%% used to denote shared contribution to the research.
% \author{Heewon Park, Mugon Joe, Miru Kim}
% \email{{heewon012, mugon, mirukim00}@soongsil.ac.kr}

% %\orcid{0009-0006-0446-3151, 0009-0003-4111-491X, 0000-0002-5394-4780}
% \affiliation{%
%   \institution{Soongsil University}
%   \state{Seoul}
%   \country{South Korea}
% }



\author{Heewon Park}
\email{heewon012@soongsil.ac.kr}
\orcid{0009-0006-0446-3151}
\authornote{H. Park and M. Joe contributed equally.}
\affiliation{%
  \institution{Soongsil University}
  \state{Seoul}
  \country{South Korea}
}
\author{Mugon Joe}
\authornotemark[1]
\email{mugon@soongsil.ac.kr}
\orcid{0009-0003-4111-491X}
\affiliation{%
  \institution{Soongsil University}
  \state{Seoul}
  \country{South Korea}
}
\author{Miru Kim}
\email{mirukim00@soongsil.ac.kr}
\orcid{0000-0002-5394-4780}
\affiliation{%
  \institution{Soongsil University}
  \state{Seoul}
  \country{South Korea}
}

\author{Minhae Kwon}
\authornote{Corresponding author.
All authors are with the Department of Intelligent Semiconductors, 
and M. Kwon is also affiliated with the School of Electronic Engineering.}
\email{minhae@ssu.ac.kr}
\affiliation{%
  \institution{Soongsil University}
  \state{Seoul}
  \country{South Korea}
}


%%
%% By default, the full list of authors will be used in the page
%% headers. Often, this list is too long, and will overlap
%% other information printed in the page headers. This command allows
%% the author to define a more concise list
%% of authors' names for this purpose.
\renewcommand{\shortauthors}{Heewon Park, Mugon Joe, Miru Kim and Minhae Kwon}

%%
%% The abstract is a short summary of the work to be presented in the
%% article.
\begin{abstract}
In real-world applications, machine learning models face \textit{online label shift}, where label distributions change over time. Effective adaptation requires careful learning rate selection: too low slows adaptation and too high causes instability. We propose \textit{ASAP} (\textbf{A}daptive \textbf{S}hift \textbf{A}ware \textbf{P}ost-training), which dynamically adjusts the learning rate by computing the cosine distance between current and previous unlabeled outputs and mapping it within a bounded range. ASAP requires no labels, model ensembles, or past inputs, using only the previous softmax output for fast, lightweight adaptation. Experiments across multiple datasets and shift scenarios show ASAP consistently improves accuracy and efficiency, making it practical for unsupervised model adaptation.
%In real-world applications, machine learning models encounter \textit{online label shift}, where the label distribution changes over time. To address this challenge, post-training with a carefully selected learning rate is required. Since a low learning rate leads to slow adaptation, and a high rate risks model instability, learning rate selection plays a critical role. In this paper, we propose \textit{ASAP (\textbf{A}daptive \textbf{S}hift \textbf{A}ware \textbf{P}ost-training)}, a simple yet effective method that dynamically adjusts the learning rate based on the estimated degree of distribution shift. Specifically, ASAP computes the cosine distance between model outputs of current and previous unlabeled data and linearly maps this distance to a learning rate within a bounded range. Unlike prior approaches, ASAP requires no labels, no ensemble of models, and no access to past input data; only the model’s previous softmax output needs to be stored. This lightweight design enables fast and robust adaptation. Extensive experiments across multiple datasets and online label shift scenarios demonstrate that ASAP consistently outperforms existing methods in both accuracy and training efficiency, making it a practical solution for real-world unsupervised model adaptation.

%Machine learning models deployed in dynamic environments often face evolving label distributions, a challenge known as \textit{online label shift}. Adapting to such shifts without access to ground-truth labels is critical for maintaining performance in real-world applications. One particularly important component in this process is the learning rate, which is often underexplored: if set too small, adaptation is slow and ineffective; if too large, it can destabilize the model. In this paper, we propose \textit{ASAP (\textbf{A}daptive \textbf{S}hift \textbf{A}ware \textbf{P}ost-training)}, a simple yet effective method that dynamically adjusts the learning rate based on distribution shift estimated from the model’s predictions. Specifically, ASAP computes the cosine distance between softmax outputs of current and previous unlabeled data and linearly maps this shift estimate to a learning rate within a bounded range. Unlike prior approaches, ASAP requires no labels, no ensemble of models, and no access to past input data; only the model’s previous softmax output needs to be stored. This lightweight design enables fast and robust adaptation. Extensive experiments across multiple datasets and online label shift scenarios demonstrate that ASAP consistently outperforms existing methods in both accuracy and training efficiency, making it a practical solution for real-world unsupervised model adaptation.
\end{abstract}

%%
%% The code below is generated by the tool at http://dl.acm.org/ccs.cfm.
%% Please copy and paste the code instead of the example below.
%%
\begin{CCSXML}
<ccs2012>
 <concept>
  <concept_id>00000000.0000000.0000000</concept_id>
  <concept_desc>Do Not Use This Code, Generate the Correct Terms for Your Paper</concept_desc>
  <concept_significance>500</concept_significance>
 </concept>
 <concept>
  <concept_id>00000000.00000000.00000000</concept_id>
  <concept_desc>Do Not Use This Code, Generate the Correct Terms for Your Paper</concept_desc>
  <concept_significance>300</concept_significance>
 </concept>
 <concept>
  <concept_id>00000000.00000000.00000000</concept_id>
  <concept_desc>Do Not Use This Code, Generate the Correct Terms for Your Paper</concept_desc>
  <concept_significance>100</concept_significance>
 </concept>
 <concept>
  <concept_id>00000000.00000000.00000000</concept_id>
  <concept_desc>Do Not Use This Code, Generate the Correct Terms for Your Paper</concept_desc>
  <concept_significance>100</concept_significance>
 </concept>
</ccs2012>
\end{CCSXML}

\ccsdesc[500]{Computing methodologies~Artificial intelligence}
\ccsdesc[500]{Computing methodologies~Unsupervised learning settings}
\ccsdesc[500]{Computing methodologies~Online learning settings}

%%
%% Keywords. The author(s) should pick words that accurately describe
%% the work being presented. Separate the keywords with commas.
\keywords{Online label shift, Unsupervised learning, Post-training}
%% A "teaser" image appears between the author and affiliation
%% information and the body of the document, and typically spans the
%% page.

% \received{20 February 2007}
% \received[revised]{12 March 2009}
% \received[accepted]{5 June 2009}

%%
%% This command processes the author and affiliation and title
%% information and builds the first part of the formatted document.
\maketitle

\section{Introduction}


\begin{figure}
\centering
\vspace{0.2cm}
\includegraphics[width=\linewidth]{Figure/overall.png} 
\hfill  
\vspace{-0.2cm}
\caption{Overview of the ASAP. ASAP estimates the degree of online label shift and assigns a learning rate accordingly: smaller shifts are assigned lower learning rates to maintain stability, while larger shifts receive higher learning rates to enhance adaptability.}
\vspace{-0.3cm}
\label{fig:overview}
\vspace{-0.3cm}
\end{figure}

Machine learning models in practical deployment settings face data streams with online label shift, where label distribution shifts occur gradually or abruptly over time~\cite{OLS1,OLS2,OLSFL2,OLSFL3,Shift-CIKM5,Shift-CIKM6,Shift-CIKM8,Shift-NeurIPS4, Shift-NeurIPS5,Shift-NeurIPS6,Shift-NeurIPS7,Shift-NeurIPS8,Shift-NeurIPS9}. Each incoming unlabeled data may exhibit changes in the underlying label distribution due to factors such as shifting user behavior, seasonal trends, or context-dependent inputs~\cite{Shift-AAAI1,Shift-AAAI2,Shift-AAAI3,Shift-AAAI4,Shift-AAAI5,Shift-CIKM3,Shift-CIKM4}. These changes can severely degrade the performance of models that were pre-trained on a static distribution, highlighting the need for continuous post-training during deployment~\cite{Shift-CIKM1,Shift-CIKM2,Shift-NeurIPS1,Shift-NeurIPS2,Shift-NeurIPS3,OLS3,OLS4,OLS5,Shift-ICML1,Shift-ICML2,Shift-ICML3,Shift-ICML4,Shift-ICML5}.
%In many real-world applications, machine learning models are deployed into environments where the data stream evolves over time~\cite{OLS1,OLS2,OLSFL2,OLSFL3,Shift-CIKM5,Shift-CIKM6,Shift-CIKM8,Shift-NeurIPS4,Shift-NeurIPS5,Shift-NeurIPS6,Shift-NeurIPS7,Shift-NeurIPS8,Shift-NeurIPS9}. Each incoming unlabeled data may exhibit changes in the underlying label distribution due to factors such as shifting user behavior, seasonal trends, or context-dependent inputs~\cite{Shift-AAAI1,Shift-AAAI2,Shift-AAAI3,Shift-AAAI4,Shift-AAAI5,Shift-CIKM3,Shift-CIKM4}. These changes can severely degrade the performance of models that were pre-trained on a static distribution, highlighting the need for continuous adaptation during deployment~\cite{Shift-CIKM1,Shift-CIKM2,Shift-NeurIPS1,Shift-NeurIPS2,Shift-NeurIPS3,OLS3,OLS4,OLS5,Shift-ICML1,Shift-ICML2,Shift-ICML3,Shift-ICML4,Shift-ICML5}.

To address this issue, unsupervised post-training methods have been developed, allowing models to update without ground-truth labels~\cite{ATLAS,FTH,Shift-AAAI1,Shift-ICML1}. These methods typically estimate the distribution of current data based on model predictions and adjust parameters using the estimated distribution. However, most of them rely on fixed learning rates or complex mechanisms such as ensembles, sliding-window memory, or historical regularization. These techniques often suffer from high computational costs and are less responsive to sudden distribution changes.
%To address this issue, unsupervised adaptation methods have been developed, allowing models to self-update without access to ground-truth labels~\cite{ATLAS,FTH,Shift-AAAI1,Shift-ICML1}. These methods typically estimate the current data distribution based on model predictions and adjust parameters accordingly. However, most of them rely on fixed learning rates or complex mechanisms such as ensembles, sliding-window memory, or historical regularization. These techniques often suffer from high computational costs and lack responsiveness to sudden distribution changes.

% 이부분 없어도 되지 않을지..?
%Notable efforts such as ATLAS~\cite{ATLAS}, FTFWH~\cite{FTH}, and UOGD~\cite{ATLAS} have shown that explicitly modeling label shift improves post-training. ATLAS addresses learning rate selection by maintaining multiple learners with different static learning rates, which are dynamically weighted according to recent performance. FTFWH stabilizes updates by averaging estimated label distributions over long or recent histories. UOGD leverages BBSE~\cite{BBSE} to infer label distribution shifts. While effective, these methods often rely on multi-model ensembles or historical memory, making them resource-intensive and harder to tune.

One key factor that significantly influences post-training performance is the choice of learning rate. It governs how quickly the model reacts to distribution changes. For example, a small learning rate results in slow adaptation to the shift, while an large learning rate leads to unstable updates and potential forgetting of useful prior knowledge. Existing methods often rely on a fixed learning rate or select from a finite candidate set~\cite{erven2011adaptive,yang2022dltta,ATLAS}, which constrains adaptability to diverse shift patterns.
%One key factor that significantly influences adaptation performance is the choice of learning rate. It governs how quickly the model reacts to distribution changes: if the rate is too small, adaptation lags behind the shift; if too large, updates become unstable and may forget useful prior knowledge. Existing methods often rely on a fixed learning rate or select from a finite candidate set~\cite{erven2011adaptive,yang2022dltta,ATLAS}, which constrains adaptability to diverse shift patterns.

To overcome these limitations, we propose a simple yet effective method, ASAP. At each timestep, ASAP adjusts the learning rate based on how much the model’s predictions change between consecutive unlabeled data stream. It quantifies this change using the cosine distance between model's outputs and maps the result to a continuous learning rate within a bounded range. This adaptive scheduling allows for smooth and responsive adjustment to distribution changes. The learning rate remains low when predictions are stable and increases in response to the magnitude of detected distribution shifts. Despite its simplicity, ASAP consistently achieves superior performance across diverse datasets and shift scenarios, while converging faster than competitive baselines.
%To overcome these limitations, we propose a simple yet effective method, ASAP. At each timestep, ASAP adjusts the learning rate based on how much the model’s predictions change between consecutive unlabeled data stream. It quantifies this change using the cosine distance between model's outputs and maps the result to a continuous learning rate within a bounded range. This fine-grained scheduling enables smooth and responsive adaptation: the learning rate remains low when predictions are stable and increases proportionally when substantial shifts are detected. Furthermore, ASAP requires no memory of past labels or predictions, no ensembles, and incurs minimal computational overhead, making it highly efficient in both time and space. Despite its simplicity, ASAP consistently achieves state-of-the-art accuracy across diverse datasets and shift scenarios, while converging faster than competitive baselines.

Our contributions are summarized as follows.
\begin{itemize}
    \item We propose a method to estimate the label distribution shift in online settings using the cosine distance between model predictions on consecutive unlabeled data samples.
    \item We propose a learning rate adjustment strategy that continuously tunes the learning rate based on the estimated magnitude of distribution shift, rather than relying on fixed or discrete candidate values.
%    \item We introduce a post-training algorithm that adjusts the learning rate based on the estimated shift, without relying on fixed or discrete candidate values.
    \item We adopt a risk estimation framework based on pseudo-label distributions to enable fully unsupervised model updates.
    \item We evaluate our method on four datasets and four types of label shift, consistently outperforming five competitive baselines in both accuracy and time efficiency.
\end{itemize}


\section{Preliminaries}
\label{preliminaries}
\vspace{-0.1cm}

\subsection{Online Label Shift}

For the pre-training phase ($t=0$), let $\theta$ denote a model trained on a pre-collected labeled data set $(\mathbf{x}^0, \mathbf{y}^0) \sim \mathbf{P}_{\mathbf{y}^0}$.  
Here, $\mathbf{P}_{\mathbf{y}^0}$ is the class distribution at time $t = 0$, where each label $\mathbf{y}^0$ is a class index $c \in \mathcal{C}$.  
The $c$-th element of $\mathbf{P}_{\mathbf{y}^0}$, denoted $[\mathbf{P}_{\mathbf{y}^0}]_c$, represents the proportion of class $c$ in the initial training labels.

At online post-training phase ($0 < t \le T$), the model encounters a sequential data stream $\{\mathbf{x}^1, \mathbf{x}^2,\dots, \mathbf{x}^T\}$, where each data $\mathbf{x}^t$ arrives without corresponding labels.  
We assume the label distribution $\mathbf{P}_{\mathbf{y}^t}$ underlying each data may shift over time, while the conditional distribution $\mathbf{P}(\mathbf{x} \mid \mathbf{y})$ remains stationary.  
This online label shift setting can be modeled using a convex interpolation as follows.
\begin{equation}
\label{eqn:label_shift_central}
\mathbf{P}_{\mathbf{y}^t} = (1 - \alpha(t))\,\mathbf{P}_{\mathbf{y}^0} + \alpha(t)\,\mathbf{P}_{\mathbf{y}^T}, \quad \alpha(t) \in [0, 1]
\end{equation}
Here, $\alpha(t)$ governs the degree of shift from the initial label distribution $\mathbf{P}_{\mathbf{y}^0}$ toward a target distribution $\mathbf{P}_{\mathbf{y}^T}$. The model $\theta$ must be adapted to reflect the evolving distribution $\mathbf{P}_{\mathbf{y}^t}$ at each time step $t$, despite lacking access to ground-truth labels.

\subsection{Learning Objective with Unsupervised Risk Estimation}
\label{section:prelim}

To support model updates with unlabeled data, we adopt an unsupervised risk estimation framework. The optimal objective with true label at time $t$ is to minimize the expected risk as follows.
%The true (but inaccessible) objective at time $t$ is to minimize the expected risk as follows.
\begin{equation}
\label{eq:risk_def}
\mathcal{R}^t(\theta) := \mathbb{E}_{(\mathbf{x}^t, \mathbf{y}^t) \sim \mathbf{P}_{\mathbf{y}^t}} \left[ \mathcal{L}( \mathcal{F}(\theta; \mathbf{x}^t), \mathbf{y}^t) \right] = \sum_{c \in \mathcal{C}} [\mathbf{P}_{\mathbf{y}^t}]_c \cdot \mathcal{R}^{t,c}(\theta)
\end{equation}
Here, $\mathcal{F}(\cdot)$ denotes the model’s softmax output, and $\mathcal{R}^{t,c}(\theta)$ represents the class-wise risk for class $c$, defined as the expected loss over inputs with label $c$.

Under the label shift assumption, class-conditional risks remain stable across time. Thus, the risk $\mathcal{R}^t(\theta)$ at time $t$ can be approximated using class-wise risks from the pre-training phase. 
\begin{equation}
\label{eq:risk_decomp_central}
\mathcal{R}^t(\theta) \approx \sum_{c \in \mathcal{C}} [\mathbf{P}_{\mathbf{y}^t}]_c \cdot \mathcal{R}^{0,c}(\theta)
\end{equation}
Since the true label distribution $\mathbf{P}_{\mathbf{y}^t}$ is unknown during online adaptation, we estimate it from the model’s soft predictions.  
Let $\mathbf{P}_{\widehat{\mathbf{y}}^t}$ be the pseudo-label distribution derived from model outputs on $\mathbf{x}^t$.  
We apply Black-box Shift Estimation (BBSE)~\cite{BBSE}, which estimates the true label distribution.
\begin{equation}
\mathbf{P}_{\mathbf{y}^t} \approx \mathbf{M}^{-1} \mathbf{P}_{\widehat{\mathbf{y}}^t}
\end{equation}
Here, $\mathbf{M}$ is the confusion matrix estimated from the initial labeled data $(\mathbf{x}^0, \mathbf{y}^0)$. Using this estimate, the risk in~\eqref{eq:risk_decomp_central} can be estimated without labels.
\begin{equation}
\label{eq:risk_estimated_central}
\widehat{\mathcal{R}}^t(\theta) := \sum_{c \in \mathcal{C}} \left[ \mathbf{M}^{-1} \mathbf{P}_{\widehat{\mathbf{y}}^t} \right]_c \cdot \mathcal{R}^{0,c}(\theta)
\end{equation}
Finally, the learning objective at time $t$ is to update model parameters by minimizing the estimated unsupervised risk.
\begin{equation}
\theta^{*} = \arg\min_{\theta} \widehat{\mathcal{R}}^t(\theta)
\end{equation}
This formulation provides a foundation for updating models in an unsupervised online setting, even in the presence of label shift.


\section{Shift Estimation-based Adaptive Post-training}
\label{section:method}
% \begin{algorithm}[t]
% \caption{ ASAP for Adaptive Post-training}
% \label{alg:adaptive_update}
% \begin{algorithmic}[1]
% \State \textbf{Input:} Pretrained model $\theta$, learning rate bounds $[\eta_{min}, \eta_{max}]$
% \For{$t = 1$ to $T$}
%     \State Estimate shift $\mathcal{E}^t$ via~\eqref{eq:shift}
%     \State Compute $\eta^t$ via~\eqref{eq:lr}
%     \State Estimate risk $\widehat{\mathcal{R}}^t(\theta)$ via~\eqref{eq:risk_estimated_central}
%     \State Update model: $\theta \leftarrow \theta - \eta^t \nabla_\theta \widehat{\mathcal{R}}^t(\theta)$
% \EndFor
% \end{algorithmic}
% \end{algorithm}

\begin{algorithm}[t]
\caption{ASAP for Adaptive Post-training}
\label{alg:adaptive_update}
\begin{algorithmic}[1]
\State \textbf{Input:} Pre-trained model $\theta$, learning rate bounds $[\eta_{\text{min}}, \eta_{\text{max}}]$
\State Initialize prediction buffer: $b^{0} \leftarrow \mathcal{F}(\theta; \mathbf{x}^0)$
\For{$t = 1$ to $T$}
    \State Receive new unlabeled data sample $\mathbf{x}^t$
    \State Compute current prediction: $b^t \leftarrow \mathcal{F}(\theta; \mathbf{x}^t)$
    \State Estimate shift $\mathcal{E}^t \leftarrow 1 - \frac{ \langle b^{t-1}, b^t \rangle }{ \|b^{t-1}\|_2 \cdot \|b^t\|_2 }$
    \State Compute learning rate: $\eta^t \leftarrow \eta_{\text{min}} + \mathcal{E}^t \cdot (\eta_{\text{max}} - \eta_{\text{min}})$
    \State Estimate unsupervised risk: $\widehat{\mathcal{R}}^t(\theta)$ via~\eqref{eq:risk_estimated_central}
    \State Update model parameters: $\theta \leftarrow \theta - \eta^t \nabla_\theta \widehat{\mathcal{R}}^t(\theta)$
    \State Update prediction buffer: $b^{t-1} \leftarrow b^t$
\EndFor
\end{algorithmic}
\end{algorithm}

We propose a shift-aware post-training framework ASAP for online learning in non-stationary environments. ASAP dynamically adjusts the model’s learning rate at each timestep according to the estimated degree of label distribution shift. The core insight is to use the change in the model output distribution, based on softmax predictions, to infer the distribution change and adjust the updates accordingly, without requiring labeled data.

\subsection{Shift-aware Learning Rate Adjustment}
As the model is deployed, it receives a sequence of unlabeled data samples $\{\mathbf{x}^1, \mathbf{x}^2, \dots, \mathbf{x}^T\}$ over time. To estimate distributional shifts between time steps, we track how much the model's predictions change between adjacent data samples.
At each time step $t$, we compute softmax predictions on the current input data samples, denoted as $b^t = \mathcal{F}(\theta; \mathbf{x}^t)$, and compare them with those from the previous data samples, stored in a prediction buffer $b^{t-1}$. The predicted shift $\mathcal{E}^t$ is defined as the cosine distance between these two distributions.
\begin{equation}
\label{eq:shift}
\mathcal{E}^t := 1 - \frac{ \left\langle b^{t-1}, b^t \right\rangle }{ | b^{t-1} |_2 \cdot | b^t |_2 }
\end{equation}
The value $\mathcal{E}^t$ lies in $[0, 1]$ when the predictions are probability distributions (e.g., softmax outputs), as the cosine distance between such vectors is bounded. A smaller $\mathcal{E}^t$ implies that the model's predictions have remained stable, suggesting little distributional change. Conversely, a larger value indicates a greater shift in the underlying data distribution.

After computing the shift estimate $\mathcal{E}^t$, we update the prediction buffer: $b^{t-1} \leftarrow b^t$, so that it can be used in the next iteration. This enables continual shift tracking without requiring storage of past data or labels. We then linearly map this shift value to a learning rate within a bounded range $[\eta_{\text{min}}, \eta_{\text{max}}]$.
\begin{equation}
\label{eq:lr}
    \eta^t = \eta_{\text{min}} + \mathcal{E}^t \cdot (\eta_{\text{max}} - \eta_{\text{min}})
\end{equation}
This design allows the learning rate $\eta^t$ to respond proportionally to the estimated magnitude of distribution shift. When the predicted distributions remain similar across consecutive timesteps (i.e., $\mathcal{E}^t$ is small), the update step is modest, helping to preserve learned knowledge and avoid unnecessary changes. In contrast, when the prediction change is large (i.e., $\mathcal{E}^t$ approaches 1), a higher learning rate is selected, allowing the model to rapidly adapt to potentially significant changes in the underlying data distribution.

%This linear scaling offers a practical and powerful approach. It avoids the need to heuristically select or tune a fixed learning rate and instead adapts it based on measurable predictive dynamics, making the model more responsive to real-time data evolution.

\subsection{Model Update with Estimated Risk}

To update the model using unlabeled data, we use the unsupervised risk estimation framework described in Section~\ref{section:prelim}. At each timestep $t$, the expected risk $\widehat{\mathcal{R}}^t(\theta)$ is estimated based on pseudo-label distributions and class-wise risk statistics from pre-training.

The model is then updated using the adaptively selected learning rate as follows.
\begin{equation}
\label{eq:model_update}
    \theta \leftarrow \theta - \eta^t \nabla_\theta \widehat{\mathcal{R}}^t(\theta)
\end{equation}
This allows the model to continuously adapt in a way that is sensitive to distribution dynamics, ensuring both responsiveness and stability across time. The procedure is summarized in Algorithm~\ref{alg:adaptive_update}.


% \section{Simulation Setup}
\section{Simulation}

In this section, we introduce the simulation setups, including the datasets, online label shift modeling, and baseline adaptation methods. Also, we conduct extensive experiments to validate the effectiveness of the ASAP.


\vspace{-0.1cm}
% \subsection{Datasets and Online Label Shift Modeling}
\subsection{Simulation Setup}
We evaluate on four standard benchmarks—Tiny ImageNet~\cite{Tinyimagenet}, CIFAR-10~\cite{CIFAR10}, FMNIST~\cite{FMNIST}, and MNIST~\cite{MNIST}. To simulate online label shift dynamics, we model the class prior evolution from an initial uniform distribution $\mathbf{P}_{\mathbf{y}^0}$ set uniformly across all classes to mirror pre-training conditions to a target Dirac delta distribution $\mathbf{P}_{\mathbf{y}^T}$ concentrated on a single class selected at random. This transition is governed by a time-dependent mixing coefficient $\alpha(t)$  as in~\eqref{eqn:label_shift_central}. Four label shift patterns are implemented through distinct $\alpha(t)$ formulations:

\begin{itemize}
    \item \textbf{Linear Shift} (Lin.)~\includegraphics[height=2ex]{Figure/linear.png}: $\alpha(t) = t/T$ produces smooth linear transitions over $T$ timesteps.
    \item \textbf{Sine Shift} (Sin.)~\includegraphics[height=2ex]{Figure/sine.png}: $\alpha(t) = \sin(\pi t/\sqrt{T})$ creates periodic oscillations in class proportions.
    \item \textbf{Square Shift} (Squ.)~\includegraphics[height=2ex]{Figure/square.png}: $\alpha(t)$ alternates between 0 and 1 every $\sqrt{T}/2$ steps, generating abrupt distribution changes.
    \item \textbf{Bernoulli Shift} (Ber.)~\includegraphics[height=2ex]{Figure/bernoulli.png}: $\alpha(t)$ flips state with probability $1/\sqrt{T}$, modeling stochastic distribution shifts.
\end{itemize}




% \subsection{Baselines Adaptation Methods}
% We compare the proposed solution with the following baselines:
% \begin{itemize}
% \item \textbf{ROGD}~\cite{FTH}: ROGD introduces a temporal regularization term to standard OGD, penalizing large deviations from previous model parameters. This regularization mitigates catastrophic forgetting and stabilizes adaptation under label shift by balancing the incorporation of new information with the retention of prior knowledge.

% \item \textbf{UOGD}~\cite{ATLAS}: UOGD extends OGD to the label shift setting by employing BBSE. It estimates the current label distribution using pseudo-labels and the pre-trained confusion matrix, enabling unbiased risk estimation and model updates without requiring access to ground-truth labels.

% \item \textbf{FTH}~\cite{FTH}: FTH maintains an average of estimated label distributions across all previous time steps. By aggregating the entire history of predictions, FTH assumes that past distribution information is informative for current adaptation, making it robust to gradual, monotonic shifts but potentially less responsive to abrupt changes.

% \item \textbf{FTFWH}~\cite{FTH}: FTFWH improves upon FTH by averaging estimated label distributions over a fixed-size sliding window of recent time steps. This approach enhances sensitivity to recent distribution changes, allowing faster adaptation to sudden shifts, with the window size controlling the trade-off between stability and reactivity.

% \item \textbf{ATLAS}~\cite{ATLAS}: ATLAS maintains an ensemble of base learners, each optimized with a distinct learning rate. Using an exponential weighting scheme, ATLAS dynamically aggregates these learners according to their recent performance, enabling robust adaptation to various patterns and magnitudes of label shift.


% \item \textbf{Proposed} (ours): Our proposed solution introduces a key point: an adaptive learning rate, computed per time step via cosine distance between consecutive softmax predictions. This adaptive learning rate enables efficient and stable adaptation to dynamic label shifts without requiring labels or costly hyperparameter tuning.
% \end{itemize}

We conduct comprehensive evaluations against five state-of-the-art algorithms for online label shift adaptation to evaluate the effectiveness of ASAP:

\begin{itemize}

\item \textbf{FTH}~\cite{FTH}: Historical averaging method that aggregates past label distributions. ASAP differs by dynamically adjusting the learning rate at each timestep, allowing for more responsive adaptation.

\item \textbf{FTFWH}~\cite{FTH}: Windowed version of FTH that focuses on recent distributions, balancing historical consistency with recent trends. ASAP instead adapts the learning rate in real time based on the estimated shift, not a fixed window size.
\item \textbf{ROGD}~\cite{FTH}: Regularized OGD with temporal constraints to prevent catastrophic forgetting. ROGD uses a fixed learning rate and does not adaptively adjust it based on distribution shift.
\item \textbf{UOGD}~\cite{ATLAS}: Unbiased OGD variant using BBSE for label distribution estimation. UOGD does not employ an adaptive learning rate mechanism and relies on a static update schedule.
\item \textbf{ATLAS}~\cite{ATLAS}: Maintains multiple base learners with different learning rates and selects the best-performing one based on recent loss. ASAP, by comparison, uses a single model and directly computes the learning rate at each time step without maintaining an ensemble.


\item \textbf{ASAP} (ours): ASAP employs cosine distance between consecutive predictions to automatically adjust learning rates, enabling stable adaptation without labels or manual tuning.
\end{itemize}


% \section{Simulation Results}

% In this section, we conduct extensive experiments to validate the effectiveness of the ASAP. We present numerical results on various datasets.

\subsection{Simulation Results}

\subsection*{Performance of Post-training}
% \begin{table}[t]
% \centering
% \caption{Performance of Post-adaptation (average accuracy (\%))}
% \label{sim:}
% \scriptsize

% \begin{tabular}{@{}cc|cccccc@{}}
% \toprule
% \textbf{Dataset} & \textbf{Shift} 
%   & ROGD    & UOGD  & FTH   & FTFWH  & ATLAS  & \textbf{Proposed} \\
% \midrule

% \multirow{4}{*}{Tiny ImageNet} 
%   & Lin. & 30.4{\tiny$\pm$1.5} & 37.3{\tiny$\pm$1.9} & 38.4{\tiny$\pm$1.6} & 39.1{\tiny$\pm$0.4} & &\textbf{42.1{\tiny$\pm$3.2}} \\
%   & Sin. & 32.4{\tiny$\pm$3.2} & 36.8{\tiny$\pm$3.0} & 37.8{\tiny$\pm$1.5} & 40.2{\tiny$\pm$0.9} & &\textbf{42.5{\tiny$\pm$2.4}} \\
%   & Squ. & 30.9{\tiny$\pm$2.7} & 34.2{\tiny$\pm$2.7} & 38.1{\tiny$\pm$1.5} & 38.3{\tiny$\pm$0.7} & &\textbf{41.4{\tiny$\pm$1.9}} \\
%   & Ber. & 28.9{\tiny$\pm$3.6} & 33.8{\tiny$\pm$2.5} & 36.3{\tiny$\pm$0.8} & 37.9{\tiny$\pm$1.3} & &\textbf{40.5{\tiny$\pm$2.7}} \\\midrule
% \multirow{4}{*}{CIFAR-10}
%   & Lin. & 62.3{\tiny$\pm$1.3} & 39.1{\tiny$\pm$1.1} & 68.3{\tiny$\pm$2.3} & 67.4{\tiny$\pm$0.2} & &\textbf{71.6{\tiny$\pm$2.4}} \\
%   & Sin. & 61.4{\tiny$\pm$1.1} & 45.5{\tiny$\pm$1.4} & 67.2{\tiny$\pm$2.4} & 65.6{\tiny$\pm$1.8} & &\textbf{70.4{\tiny$\pm$2.2}} \\
%   & Squ. & 53.8{\tiny$\pm$0.8} & 38.3{\tiny$\pm$1.1} & 62.6{\tiny$\pm$2.6} & 59.4{\tiny$\pm$2.2} & &\textbf{66.3{\tiny$\pm$2.6}} \\
%   & Ber. & 53.5{\tiny$\pm$0.6} & 38.6{\tiny$\pm$0.9} & 62.9{\tiny$\pm$2.6} & 59.3{\tiny$\pm$2.6} & &\textbf{66.6{\tiny$\pm$2.4}} \\
% \midrule
% \multirow{4}{*}{FMNIST} 
%   & Lin. & 80.3{\tiny$\pm$2.4} & 75.2{\tiny$\pm$4.6} & 84.4{\tiny$\pm$0.6} & 85.1{\tiny$\pm$0.4} & &\textbf{85.4{\tiny$\pm$0.4}} \\
%   & Sin. & 81.2{\tiny$\pm$0.3} & 74.3{\tiny$\pm$5.3} & 83.9{\tiny$\pm$0.9} & 84.0{\tiny$\pm$1.1} & &\textbf{85.3{\tiny$\pm$0.2}} \\
%   & Squ. & 79.1{\tiny$\pm$1.0} & 67.9{\tiny$\pm$5.3} & 82.5{\tiny$\pm$0.9} & 83.3{\tiny$\pm$1.7} & &\textbf{84.7{\tiny$\pm$0.8}} \\
%   & Ber. & 77.7{\tiny$\pm$1.3} & 68.2{\tiny$\pm$8.5} & 81.4{\tiny$\pm$0.9} & 82.0{\tiny$\pm$1.6} & &\textbf{84.1{\tiny$\pm$0.8}} \\\midrule
% \multirow{4}{*}{MNIST} 
%    & Lin. & 96.7{\tiny$\pm$0.4} & 96.4{\tiny$\pm$0.6} & 97.3{\tiny$\pm$0.5} & 97.9{\tiny$\pm$0.4} & &\textbf{98.8{\tiny$\pm$0.4}} \\
%    & Sin. & 96.6{\tiny$\pm$0.3} & 96.5{\tiny$\pm$0.4} & 97.2{\tiny$\pm$0.4} & 98.1{\tiny$\pm$0.5} & &\textbf{98.6{\tiny$\pm$0.2}} \\
%    & Squ. & 96.5{\tiny$\pm$0.4} & 96.5{\tiny$\pm$0.2} & 97.2{\tiny$\pm$0.1} & 97.4{\tiny$\pm$0.5} & &\textbf{98.5{\tiny$\pm$0.5}} \\
%    & Ber. & 97.3{\tiny$\pm$0.5} & 96.5{\tiny$\pm$0.3} & 97.3{\tiny$\pm$0.7} & 97.1{\tiny$\pm$1.0} & &\textbf{98.0{\tiny$\pm$0.4}} \\
% \bottomrule
% \end{tabular}
% \end{table}



\begin{table}[t]
\centering
\caption{Performance of post-training (average accuracy (\%))}
\label{sim:results}
\small
\setlength{\tabcolsep}{2pt}
\resizebox{\linewidth}{!}{%
\begin{tabular}{@{}cc|ccccc>{\columncolor{gray!15}}c@{}}
\toprule
\textbf{Dataset} & \textbf{Shift} 
  & FTH & FTFWH & ROGD & UOGD & ATLAS & \textbf{ASAP} (ours) \\
\midrule

\multirow{4}{*}{Tiny ImageNet} 
  & Lin. & 69.0{\tiny$\pm$13.5} & 23.6{\tiny$\pm$6.62} & 45.8{\tiny$\pm$9.8} & 45.8{\tiny$\pm$9.8} & 58.7{\tiny$\pm$10.0} & \textbf{72.4{\tiny$\pm$0.5}} \\
  & Sin. & 42.2{\tiny$\pm$10.0} & 50.7{\tiny$\pm$14.9} & 47.4{\tiny$\pm$10.0} & 47.4{\tiny$\pm$10.0} & 56.5{\tiny$\pm$10.5} & \textbf{72.3{\tiny$\pm$1.1}} \\
  & Squ. & 44.7{\tiny$\pm$10.3} & 51.9{\tiny$\pm$0.1} & 49.8{\tiny$\pm$10.6} & 49.8{\tiny$\pm$10.6} & 57.5{\tiny$\pm$11.4} & \textbf{72.2{\tiny$\pm$0.2}} \\
  & Ber. & 42.7{\tiny$\pm$12.3} & 53.6{\tiny$\pm$7.6} & 47.9{\tiny$\pm$12.8} & 47.9{\tiny$\pm$12.8} & 55.6{\tiny$\pm$13.8} & \textbf{72.5{\tiny$\pm$0.7}} \\
\midrule

\multirow{4}{*}{CIFAR-10}
  & Lin. & 66.4{\tiny$\pm$8.0} & 66.4{\tiny$\pm$8.1} & 51.1{\tiny$\pm$23.1} & 73.9{\tiny$\pm$9.9} & 60.7{\tiny$\pm$1.4} & \textbf{77.4{\tiny$\pm$3.4}} \\
  & Sin. & 51.6{\tiny$\pm$0.6} & 52.8{\tiny$\pm$1.3} & 48.7{\tiny$\pm$5.5} & 48.6{\tiny$\pm$0.8} & 60.1{\tiny$\pm$2.8} & \textbf{77.0{\tiny$\pm$3.2}} \\
  & Squ. & 42.8{\tiny$\pm$2.6} & 43.1{\tiny$\pm$2.4} & 32.7{\tiny$\pm$10.1} & 25.3{\tiny$\pm$6.2} & 61.3{\tiny$\pm$2.5} & \textbf{77.7{\tiny$\pm$3.4}} \\
  & Ber. & 19.5{\tiny$\pm$3.3} & 22.3{\tiny$\pm$4.9} & 23.5{\tiny$\pm$19.3} & 22.3{\tiny$\pm$6.7} & 59.7{\tiny$\pm$3.5} & \textbf{76.5{\tiny$\pm$2.2}} \\
\midrule

\multirow{4}{*}{FMNIST} 
  & Lin. & 71.1{\tiny$\pm$25.4} & 33.7{\tiny$\pm$10.4} & 63.1{\tiny$\pm$30.8} & 53.4{\tiny$\pm$28.3} & 86.4{\tiny$\pm$0.3} & \textbf{88.8{\tiny$\pm$0.7}} \\
  & Sin. & 71.0{\tiny$\pm$5.7} & 59.7{\tiny$\pm$19.5} & 83.0{\tiny$\pm$2.1} & 81.8{\tiny$\pm$2.5} & 86.5{\tiny$\pm$0.6} & \textbf{88.3{\tiny$\pm$0.9}} \\
  & Squ. & 75.3{\tiny$\pm$2.7} & 65.3{\tiny$\pm$11.8} & 59.8{\tiny$\pm$30.2} & 58.1{\tiny$\pm$29.7} & 86.3{\tiny$\pm$0.5} & \textbf{88.7{\tiny$\pm$0.9}} \\
  & Ber. & 74.8{\tiny$\pm$2.1} & 64.9{\tiny$\pm$13.7} & 82.2{\tiny$\pm$1.6} & 73.5{\tiny$\pm$9.1} & 86.0{\tiny$\pm$0.5} & \textbf{88.9{\tiny$\pm$1.0}} \\
\midrule

\multirow{4}{*}{MNIST} 
  & Lin. & 84.6{\tiny$\pm$24.6} & 96.5{\tiny$\pm$1.1} & 63.0{\tiny$\pm$17.7} & 81.1{\tiny$\pm$19.9} & 98.8{\tiny$\pm$0.2} & \textbf{99.2{\tiny$\pm$0.1}} \\
  & Sin. & 87.9{\tiny$\pm$7.3} & 85.5{\tiny$\pm$15.0} & 88.4{\tiny$\pm$9.3} & 77.2{\tiny$\pm$14.6} & 98.2{\tiny$\pm$0.3} & \textbf{99.2{\tiny$\pm$0.1}} \\
  & Squ. & 90.8{\tiny$\pm$6.1} & 82.5{\tiny$\pm$17.3} & 92.7{\tiny$\pm$5.1} & 83.5{\tiny$\pm$7.2} & 97.6{\tiny$\pm$0.5} & \textbf{99.2{\tiny$\pm$0.2}} \\
  & Ber. & 92.6{\tiny$\pm$5.5} & 93.4{\tiny$\pm$6.5} & 79.6{\tiny$\pm$17.5} & 76.8{\tiny$\pm$19.5} & 97.8{\tiny$\pm$0.5} & \textbf{99.2{\tiny$\pm$0.2}} \\

\bottomrule
\end{tabular}%
} \vspace{-0.5cm}
\end{table}

Table~\ref{sim:results} shows the average adaptation accuracy and standard deviation of seven online label shift adaptation methods across four datasets and four shift types, averaged over five random seeds. The wall time for each method is measured in seconds and averaged across all shift types. The ASAP consistently outperforms all baselines across all simulation results while maintaining competitive computational efficiency.

Across all datasets and shift settings, our proposed ASAP achieves the average relative improvement rates of  20.8\%, compared to the next best-performing baseline method. Furthermore, ASAP maintains stable and superior performance, with a low standard deviation of 4.27, computed across four shift types for each dataset. This is in contrast to the baselines, which suffer from highly inconsistent performance across different datasets and shift types, resulting in an average standard deviation of 13.8 across all methods. This improvement is attributed to the use of an adaptive learning rate, enabling more precise and efficient adaptation to evolving label distributions. These results demonstrate the effectiveness of the ASAP in handling online label shifts.


% On Tiny ImageNet, the ASAP achieves an average accuracy of 87.0\%, whereas the next best baseline, ATLAS, records an average of 51.1\%. For CIFAR-10, the ASAP achieves an average accuracy of 77.0\%, while the ATLAS has an average of 60.6\%. On FMNIST and MNIST, most methods achieve relatively high accuracy, but the ASAP still records the best results, reaching up to 88.8\% on FMNIST and around 99.2\% on MNIST.  Across all shift types, the ASAP maintains stable and superior performance, while other baselines often show significant drops depending on the dataset or shift pattern. This improvement is attributed to the use of a label-shift-adaptive learning rate that is dynamically adjusted at each timestep, enabling more precise and efficient adaptation to evolving label distributions. These results demonstrate the effectiveness of the ASAP in handling online label shifts.

In computational efficiency, the wall time comparison in Table~\ref{tab:walltime} reveals that ASAP achieves superior computational efficiency across all datasets. ASAP consistently records the lowest wall time per timestep, with an average improvement of 20.3\% in adaptation speed compared to the next-best baseline. This efficiency advantage is derived from directly computing the learning rate through mathematical formulation rather than maintaining multiple candidate models. Unlike ATLAS, which requires evaluating numerous base learners with different learning rates, ASAP adapts a single model. This eliminates the computational overhead associated with ensemble maintenance and model selection, making ASAP particularly suitable for resource-constrained environments where both adaptation performance and computational efficiency are essential.

%\vspace{-0.2cm}
\subsection*{Parameter Sensitivity Analysis}
We analyze the sensitivity of ASAP to its two key hyperparameters, the minimum and maximum learning rates (\(\eta_{\text{min}}\), \(\eta_{\text{max}}\)), in Figure~\ref{fig:param}. 
For each parameter, we vary its value while keeping the other fixed: \(\eta_{\text{max}} = 1 \times 10^{-4}\) during the \(\eta_{\text{min}}\) analysis, and \(\eta_{\text{min}} = 5 \times 10^{-6}\) for the \(\eta_{\text{max}}\) analysis. 
Each plot reports average accuracy across four datasets and four label shift types, with error bars representing one standard deviation over five random seeds.

In Figure~\ref{fig:param}(a), we observe that extremely small values of \(\eta_{\text{min}} < 1 \times 10^{-6}\) lead to suboptimal adaptation, as the model fails to adjust even when the distribution changes. Conversely, large values approaching \(\eta_{\text{max}}\) (e.g., \(5 \times 10^{-5}\)) limit the modulation range of the adaptive scheduler, resulting in unstable or overly reactive updates. 
Stable and high performance is consistently observed when \(\eta_{\text{min}}\) is selected within the range \([1 \times 10^{-6}, 1 \times 10^{-5}]\).

In Figure~\ref{fig:param}(b), when analyzing \(\eta_{\text{max}}\), we find that values larger than \(1 \times 10^{-3}\) cause accuracy to deteriorate sharply. 
Setting \(\eta_{\text{max}} = 1 \times 10^{-2}\) causes an average accuracy drop of \(11.47\%\) compared to the optimal value at \(1 \times 10^{-4}\), due to overly large gradient steps destabilizing the model. 
In contrast, when \(\eta_{\text{max}}\) lies within the moderate range of \([1 \times 10^{-5}, 1 \times 10^{-3}]\), ASAP remains stable and achieves high accuracy across all datasets. 
%These results confirm that ASAP is robust across a wide range of reasonable settings, provided extreme learning rate values are avoided.
These results confirm that ASAP is robust across a wide range of reasonable settings, as long as extreme learning rate values are avoided

%Figure~\ref{fig:param} shows the sensitivity analysis for the minimum and maximum learning rate parameters, $\eta_\text{min}$ and $\eta_\text{max}$. Each figure contains results for all four datasets and averages over four shift types. The x-axis represents the parameter values and the y-axis shows the post-training accuracy. Error bars indicate one standard deviation computed from five random seeds.

%The results indicate that the method maintains stable performance across a broad range of parameter values. When $\eta_\text{max}$ is set between $5 \times 10^{-5}$ and $5 \times 10^{-3}$ and $\eta_\text{min}$ is set between $1 \times 10^{-6}$ and $1 \times 10^{-4}$, the method achieves accuracy close to its peak value. Even when the parameters are set to more extreme values, the decrease in accuracy is minor compared to baseline methods. This robustness across datasets and shift types demonstrates that the method is not sensitive to the choice of these parameters, making it practical to use without extensive hyperparameter tuning.
\begin{table}[t]
\centering
\caption{Wall time (sec.) of each method on four datasets}
\label{tab:walltime}
\small
\setlength{\tabcolsep}{2pt}
\resizebox{\linewidth}{!}{%
\begin{tabular}{@{}c|ccccc>{\columncolor{gray!15}}c@{}}
\toprule
\textbf{Dataset} 
  & FTH & FTFWH & ROGD & UOGD & ATLAS & \textbf{ASAP} (ours) \\
\midrule

Tiny ImageNet 
  & 15.98{\tiny$\pm$0.07} & 15.43{\tiny$\pm$0.01} & 44.09{\tiny$\pm$0.11} & 44.12{\tiny$\pm$0.05} & 48.11{\tiny$\pm$0.02} & \textbf{14.54{\tiny$\pm$0.03}} \\
\midrule

CIFAR-10
  & 1.58{\tiny$\pm$0.09} & 1.62{\tiny$\pm$0.11} & 2.45{\tiny$\pm$0.12} & 2.33{\tiny$\pm$0.15} & 3.29{\tiny$\pm$0.08} & \textbf{1.12{\tiny$\pm$0.07}} \\
\midrule

FMNIST 
  & 1.52{\tiny$\pm$0.06} & 1.50{\tiny$\pm$0.07} & 2.37{\tiny$\pm$0.05} & 2.23{\tiny$\pm$0.15} & 4.20{\tiny$\pm$0.08} & \textbf{1.13{\tiny$\pm$0.09}} \\
\midrule

MNIST 
  & 1.44{\tiny$\pm$0.11} & 1.58{\tiny$\pm$0.11} & 2.29{\tiny$\pm$0.18} & 2.03{\tiny$\pm$0.14} & 7.26{\tiny$\pm$0.34} & \textbf{1.13{\tiny$\pm$0.31}} \\
\bottomrule
\end{tabular}%
}\vspace{-0.3cm}
\end{table}

\begin{figure}[t]
    \centering
    \vspace{-0.1cm}
    \begin{subfigure}[b]{0.4\textwidth}
        \includegraphics[width=\linewidth]{Figure/legend_only.png}
        % \caption*{}
    \end{subfigure}
    
    \vspace{-0.2cm}

    \begin{subfigure}[b]{0.22\textwidth}
        \includegraphics[width=\linewidth]{Figure/plot_lrmin.png}
         \vspace{-0.7cm}
        \caption{ Performance under varying \(\eta_{\text{min}}\) with fixed \(\eta_{\text{max}} = 1 \times 10^{-4}\) }
    \end{subfigure}
    % \hspace{1em}
    \begin{subfigure}[b]{0.22\textwidth}
        \includegraphics[width=\linewidth]{Figure/plot_lrmax.png}
         \vspace{-0.7cm}
        \caption{Performance under varying \(\eta_{\text{max}}\) with fixed \(\eta_{\text{min}} = 5 \times 10^{-6}\)}
        %\vspace{-0.3cm}
    \end{subfigure}
    
    \caption{Sensitivity analysis of learning rate parameters}
    \label{fig:param}
    %\vspace{-0.2cm}
\end{figure}
\vspace{-0.1cm}
\subsection*{Learning Rate Selection Analysis}

\begin{wrapfigure}{r}{0.25\textwidth}
    \centering
    \vspace{-0.5cm}
    \includegraphics[width=\linewidth]{Figure/selected_lr.png}
    \caption{Selected learning rate according to predicted shift $\mathcal{E}^t$ on Tiny ImageNet}
    \label{fig:estimation}
    \vspace{-0.5cm}
\end{wrapfigure}
% Figure~\ref{fig:tiny_results} illustrates the post-adaptation loss and accuracy curves. Each figure reports results over 100 timesteps under four types of label shift. The top row in each figure shows the loss, which decreases for all methods, indicating that model adaptation is taking place. However, in this unsupervised setting, a lower loss does not necessarily correspond to improved classification accuracy. The bottom row presents the classification accuracy, where the ASAP maintains stable and high performance throughout the adaptation period, consistently outperforming the baselines. In particular, under sinusoidal and square shift scenarios, all methods display oscillatory accuracy patterns that reflect the underlying label distribution changes, but the ASAP achieves higher average accuracy and reduced performance fluctuations compared to other methods. These results confirm that the proposed approach is effective in maintaining robust classification performance under various non-stationary shift patterns.

% Figure~\ref{fig:estimation} shows the relationship between the estimated label shift magnitude on the x-axis and the learning rate chosen by the algorithm on the y-axis.
We visualize how the learning rate is selected at levels of estimated label shift in Figure~\ref{fig:estimation}, where the x-axis represents the shift magnitude and the y-axis denotes the chosen learning rate. The results indicate that when the estimated shift is large, the algorithm tends to select a larger learning rate, while a smaller learning rate is chosen when the estimated shift is small. 

The selected learning rate also differs according to the type of label shift. Lin. produces gradual changes in the learning rate, while Sin. results in periodic fluctuations. Squ. and Ber. lead to irregular learning rates during abrupt changes. These results demonstrate that the method adaptively adjusts the learning rate according to the underlying shift pattern. This adaptive behavior indicates that the ASAP can effectively adjust its learning rate in response to the degree of distributional change, supporting efficient adaptation to rapidly changing environments while maintaining stability when the distribution is relatively constant.

%\vspace{-0.1cm}
\section{Conclusion}
% We proposed a novel unsupervised online adaptation framework under label distribution shift. ASAP utilizes a shift-aware adaptive learning rate, per-timestep adaptation without true labels. Empirical results across multiple datasets and shift types confirm that the ASAP achieves state-of-the-art adaptation performance. As a broader impact, our method could help improve robustness in real-world environments where label distribution changes over time.
%We proposed a novel unsupervised online post-training method under label distribution shift, where the model update is driven by a shift-aware adaptive learning rate computed at each timestep without requiring true labels. The algorithm dynamically adjusts its learning rate in response to the estimated magnitude of label shift, allowing for rapid and stable adaptation to changing distributions. Simulation results across multiple datasets and shift types confirm that ASAP achieves superior performance compared to existing methods. The method is robust across a wide range of parameter settings and consistently maintains high accuracy even under challenging and rapidly changing label shift scenarios. 
We proposed an unsupervised online post-training method for label distribution shift, using a shift-aware adaptive learning rate computed at each timestep without true labels. By adjusting the rate to the estimated shift magnitude, the model adapts rapidly and stably to changing distributions. Experiments across datasets and shift types show ASAP delivers superior and robust performance, maintaining high accuracy even under challenging, rapidly changing scenarios.
% As a broader impact, our method could help improve robustness in real-world environments where label distribution changes over time.

\section*{Acknowledgment}
This research was supported in part by National Research Foundation of Korea (NRF) grant (RS-2023-00278812, RS-2025-02214082), and in part by the Institute of Information \& communications Technology Planning \& Evaluation (IITP) grants (IITP-2025-RS-2020-II201602) funded by the Korea government (MSIT).

\newpage

\section*{GenAI Usage Disclosure}

We acknowledge the use of Generative AI (GenAI) tools in the preparation of this paper as follows:

\begin{itemize}
    \item \textbf{Writing assistance:} ChatGPT (OpenAI) was used for improving grammar, rephrasing, and refining the clarity of certain paragraphs. All substantive content and structure were authored by the authors.
    \item \textbf{Code generation:} No GenAI tools were used to generate or write code used in this study.
    \item \textbf{Data processing or analysis:} No GenAI tools were used for data processing, analysis, or result generation.
\end{itemize}

All uses of GenAI tools complied with the ACM Authorship Policy on Generative AI usage.


\bibliographystyle{unsrt}
%\bibliographystyle{ACM-Reference-Format}
\bibliography{reference}


\end{document}
\endinput
%%
%% End of file `sample-sigconf.tex'.

\begin{thebibliography}{10}

\bibitem{OLS1}
YuYang Qian, Yong Bai, ZhenYu Zhang, Peng Zhao, and ZhiHua Zhou.
\newblock {Handling new class in online label shift}.
\newblock In {\em IEEE International Conference on Data Mining}, 2023.

\bibitem{OLS2}
Sunghyun Park, Seunghan Yang, Jaegul Choo, and Sungrack Yun.
\newblock {Label shift adapter for test-time adaptation under covariate and label shifts}.
\newblock In {\em IEEE/CVF International Conference on Computer Vision}, 2023.

\bibitem{OLSFL2}
Zhenheng Tang, Yonggang Zhang, Peijie Dong, Yiu-ming Cheung, Amelie Zhou, Bo~Han, and Xiaowen Chu.
\newblock {FuseFL: One-shot federated learning through the lens of causality with progressive model fusion}.
\newblock In {\em Advances in Neural Information Processing Systems}, 2024.

\bibitem{OLSFL3}
Naibo Wang, Yuchen Deng, Wenjie Feng, Shichen Fan, Jianwei Yin, and See-Kiong Ng.
\newblock One-shot sequential federated learning for non-iid data by enhancing local model diversity.
\newblock In {\em Proceedings of the ACM International Conference on Multimedia}, 2024.

\bibitem{Shift-CIKM5}
Hongyan Hao, Zhixuan Chu, Shiyi Zhu, Gangwei Jiang, Yan Wang, Caigao Jiang, James~Y Zhang, Wei Jiang, Siqiao Xue, and Jun Zhou.
\newblock Continual learning in predictive autoscaling.
\newblock In {\em Proceedings of the ACM International Conference on Information and Knowledge Management}, 2023.

\bibitem{Shift-CIKM6}
Kai Yao, Zixian Su, Xi~Yang, Jie Sun, and Kaizhu Huang.
\newblock Explore epistemic uncertainty in domain adaptive semantic segmentation.
\newblock In {\em Proceedings of the ACM International Conference on Information and Knowledge Management}, 2023.

\bibitem{Shift-CIKM8}
Jie Liao, Jintang Li, Liang Chen, Bingzhe Wu, Yatao Bian, and Zibin Zheng.
\newblock {SAILOR}: Structural augmentation based tail node representation learning.
\newblock In {\em Proceedings of the ACM International Conference on Information and Knowledge Management}, 2023.

\bibitem{Shift-NeurIPS4}
Xingyu Zhu, Beier Zhu, Yi~Tan, Shuo Wang, Yanbin Hao, and Hanwang Zhang.
\newblock Enhancing zero-shot vision models by label-free prompt distribution learning and bias correcting.
\newblock In {\em Advances in Neural Information Processing Systems}, 2024.

\bibitem{Shift-NeurIPS5}
Pratiksha Thaker, Amrith Setlur, Zhiwei~S Wu, and Virginia Smith.
\newblock On the benefits of public representations for private transfer learning under distribution shift.
\newblock In {\em Advances in Neural Information Processing Systems}, 2024.

\bibitem{Shift-NeurIPS6}
Zhen-Yu Zhang, Zhiyu Xie, Huaxiu Yao, and Masashi Sugiyama.
\newblock Test-time adaptation in non-stationary environments via adaptive representation alignment.
\newblock In {\em Advances in Neural Information Processing Systems}, 2024.

\bibitem{Shift-NeurIPS7}
Yarin Bar, Shalev Shaer, and Yaniv Romano.
\newblock Protected test-time adaptation via online entropy matching: A betting approach.
\newblock In {\em Advances in Neural Information Processing Systems}, 2024.

\bibitem{Shift-NeurIPS8}
Eungyeup Kim, Mingjie Sun, Christina Baek, Aditi Raghunathan, and J~Zico Kolter.
\newblock Test-time adaptation induces stronger accuracy and agreement-on-the-line.
\newblock In {\em Advances in Neural Information Processing Systems}, 2024.

\bibitem{Shift-NeurIPS9}
Yoonki Cho, Jaeyoon Kim, Woo~J Kim, Junsik Jung, and Sung-eui Yoon.
\newblock Generalizable person re-identification via balancing alignment and uniformity.
\newblock In {\em Advances in Neural Information Processing Systems}, 2024.

\bibitem{Shift-AAAI1}
Shreyas Havaldar, Jatin Chauhan, Karthikeyan Shanmugam, Jay Nandy, and Aravindan Raghuveer.
\newblock Fairness under covariate shift: Improving fairness-accuracy tradeoff with few unlabeled test samples.
\newblock In {\em Proceedings of the AAAI Conference on Artificial Intelligence}, 2024.

\bibitem{Shift-AAAI2}
Anh~T Nguyen, Lam Tran, Anh Tong, Tuan-Duy~H Nguyen, and Toan Tran.
\newblock {CASUA}l: Conditional support alignment for domain adaptation with label shift.
\newblock In {\em Proceedings of the AAAI Conference on Artificial Intelligence}, 2025.

\bibitem{Shift-AAAI3}
Dexter Neo, Stefan Winkler, and Tsuhan Chen.
\newblock {MaxEnt} loss: constrained maximum entropy for calibration under out-of-distribution shift.
\newblock In {\em Proceedings of the AAAI Conference on Artificial Intelligence}, 2024.

\bibitem{Shift-AAAI4}
Yujie Chen, Wenhui Wu, Le~Ou-Yang, Ran Wang, and Debby~D Wang.
\newblock {GeCC}: Generalized contrastive clustering with domain shifts modeling.
\newblock In {\em Proceedings of the AAAI Conference on Artificial Intelligence}, 2025.

\bibitem{Shift-AAAI5}
Yanan Wu, Zhixiang Chi, Yang Wang, Konstantinos~N Plataniotis, and Songhe Feng.
\newblock Test-time domain adaptation by learning domain-aware batch normalization.
\newblock In {\em Proceedings of the AAAI Conference on Artificial Intelligence}, 2024.

\bibitem{Shift-CIKM3}
Guoxin Chen, Yongqing Wang, Fangda Guo, Qinglang Guo, Jiangli Shao, Huawei Shen, and Xueqi Cheng.
\newblock Causality and independence enhancement for biased node classification.
\newblock In {\em Proceedings of the ACM International Conference on Information and Knowledge Management}, 2023.

\bibitem{Shift-CIKM4}
Anique Tahir, Lu~Cheng, and Huan Liu.
\newblock Fairness through aleatoric uncertainty.
\newblock In {\em Proceedings of the ACM International Conference on Information and Knowledge Management}, 2023.

\bibitem{Shift-CIKM1}
Dong Li, Chen Zhao, Minglai Shao, and Wenjun Wang.
\newblock Learning fair invariant representations under covariate and correlation shifts simultaneously.
\newblock In {\em Proceedings of the ACM International Conference on Information and Knowledge Management}, 2024.

\bibitem{Shift-CIKM2}
Naman Khetan, Sanyog Dewani, Gokul Swamy, and Vikalp Gajbhiye.
\newblock {XCapsUTL}: Cross-domain unsupervised transfer learning framework using a capsule neural network.
\newblock In {\em Proceedings of thed ACM International Conference on Information and Knowledge Management}, 2024.

\bibitem{Shift-NeurIPS1}
Yuli Slavutsky and Yuval Benjamini.
\newblock Class distribution shifts in zero-shot learning: Learning robust representations.
\newblock In {\em Advances in Neural Information Processing Systems}, 2024.

\bibitem{Shift-NeurIPS2}
Surbhi Goel, Abhishek Shetty, Konstantinos Stavropoulos, and Arsen Vasilyan.
\newblock Tolerant algorithms for learning with arbitrary covariate shift.
\newblock In {\em Advances in Neural Information Processing Systems}, 2024.

\bibitem{Shift-NeurIPS3}
Jiayun Wu, Jiashuo Liu, Peng Cui, and Steven~Z Wu.
\newblock Bridging multicalibration and out-of-distribution generalization beyond covariate shift.
\newblock In {\em Advances in Neural Information Processing Systems}, 2024.

\bibitem{OLS3}
Chuang Chen, Jiantao Shi, Mouquan Shen, Ningyun Lu, Hui Yu, Yukun Chen, and Cunsong Wang.
\newblock {Pseudo-label guided sparse deep belief network learning method for fault diagnosis of radar critical components}.
\newblock {\em IEEE Transactions on Instrumentation and Measurement}, 72:1--12, 2023.

\bibitem{OLS4}
Suresh Amalapuram, Bheemarjuna Tamma, and Sumohana Channappayya.
\newblock {SPIDER: A semi-supervised continual learning-based network intrusion detection system}.
\newblock In {\em IEEE Conference on Computer Communications}, 2024.

\bibitem{OLS5}
Dheeraj Baby, Saurabh Garg, TzuChing Yen, Sivaraman Balakrishnan, Zachary Lipton, and YuXiang Wang.
\newblock {Online label shift: Optimal dynamic regret meets practical algorithms}.
\newblock In {\em Advances in Neural Information Processing Systems}, 2024.

\bibitem{Shift-ICML1}
Yue He, Dongbai Li, Pengfei Tian, Han Yu, Jiashuo Liu, Hao Zou, and Peng Cui.
\newblock Domain-wise data acquisition to improve performance under distribution shift.
\newblock In {\em International Conference on Machine Learning}, 2024.

\bibitem{Shift-ICML2}
Tong Wei, Zhen Mao, Zi-Hao Zhou, Yuanyu Wan, and Min-Ling Zhang.
\newblock Learning label shift correction for test-agnostic long-tailed recognition.
\newblock In {\em International Conference on Machine Learning}, 2024.

\bibitem{Shift-ICML3}
Jae-Hong Lee and Joon-Hyuk Chang.
\newblock Stationary latent weight inference for unreliable observations from online test-time adaptation.
\newblock In {\em International Conference on Machine Learning}, 2024.

\bibitem{Shift-ICML4}
Dapeng Hu, Jian Liang, Xinchao Wang, and Chuan-Sheng Foo.
\newblock Pseudo-calibration: improving predictive uncertainty estimation in unsupervised domain adaptation.
\newblock In {\em International Conference on Machine Learning}, 2024.

\bibitem{Shift-ICML5}
Wei Wang, Chao Huang, Jie Wen, and Cong Wang.
\newblock Batch singular value polarization and weighted semantic augmentation for universal domain adaptation.
\newblock In {\em International Conference on Machine Learning}, 2024.

\bibitem{ATLAS}
Yong Bai, YuJie Zhang, Peng Zhao, Masashi Sugiyama, and ZhiHua Zhou.
\newblock {Adapting to online label shift with provable guarantees}.
\newblock In {\em Advances in Neural Information Processing Systems}, 2022.

\bibitem{FTH}
Ruihan Wu, Chuan Guo, Yi~Su, and Kilian Weinberger.
\newblock {Online adaptation to label distribution shift}.
\newblock In {\em Advances in Neural Information Processing Systems}, 2021.

\bibitem{erven2011adaptive}
Tim Erven, Wouter~M Koolen, Steven Rooij, and Peter Gr{\"u}nwald.
\newblock Adaptive hedge.
\newblock {\em Advances in Neural Information Processing Systems}, 2011.

\bibitem{yang2022dltta}
Hongzheng Yang, Cheng Chen, Meirui Jiang, Quande Liu, Jianfeng Cao, Pheng~Ann Heng, and Qi~Dou.
\newblock {DLTTA: Dynamic learning rate for test-time adaptation on cross-domain medical images}.
\newblock {\em IEEE Transactions on Medical Imaging}, 41(12):3575--3586, 2022.

\bibitem{BBSE}
Zachary Lipton, Yu-Xiang Wang, and Alexander Smola.
\newblock Detecting and correcting for label shift with black box predictors.
\newblock In {\em International conference on machine learning}, 2018.

\bibitem{Tinyimagenet}
Yann Le and Xuan Yang.
\newblock Tiny imagenet visual recognition challenge.
\newblock {\em CS 231N}, 7(7):3, 2015.

\bibitem{CIFAR10}
Alex Krizhevsky and Geoffrey Hinton.
\newblock Learning multiple layers of features from tiny images.
\newblock {\em Technical report, Univ. of Toronto}, 2009.

\bibitem{FMNIST}
Han Xiao, Kashif Rasul, and Roland Vollgraf.
\newblock {Fashion-MNIST: A novel image dataset for benchmarking machine learning algorithms}.
\newblock {\em arXiv preprint arXiv:1708.07747}, 2017.

\bibitem{MNIST}
Yann LeCun, Bernhard Boser, John~S Denker, Donnie Henderson, Richard~E Howard, Wayne Hubbard, and Lawrence~D Jackel.
\newblock Backpropagation applied to handwritten zip code recognition.
\newblock {\em Neural Computation}, 1(4):541--551, 1989.

\end{thebibliography}