E2PL: Effective and Efficient Prompt Learning for Incomplete Multi-view Multi-Label Class Incremental Learning
Extracted source text · arXiv 2601.17076
\documentclass[sigconf]{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{2026}
\acmYear{2026}
\acmDOI{10.1145/3774904.3792572}
%% These commands are for a PROCEEDINGS abstract or paper.
\acmConference[WWW '26]{the ACM WEB Conference}{April 13--17,
2026}{Dubai, United Arab Emirates}
%%
%% 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.
\usepackage{multirow}
\usepackage{algorithm}
\usepackage{algorithmic}
\begin{document}
%%
%% The "title" command has an optional parameter,
%% allowing the author to define a "short title" to be used in page headers.
\title{E2PL: Effective and Efficient Prompt Learning for Incomplete Multi-view
Multi-Label Class Incremental Learning}
%%
%% 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{Jiajun Chen}
\affiliation{%
\institution{Zhejiang University}
\department{Center for Data Science}
\city{Hangzhou}
\state{Zhejiang}
\country{China}
}
\email{chenjjcccc@zju.edu.cn}
\author{Yue Wu}
\affiliation{%
\institution{Ant Group}
\city{Hangzhou}
\state{Zhejiang}
\country{China}
}
\email{yuyue.wy@antgroup.com}
\author{Kai Huang}
\affiliation{%
\institution{Ant Group}
\city{Hangzhou}
\state{Zhejiang}
\country{China}
}
\email{kevin.hk@antgroup.com}
\author{Wenxi Zhao}
\affiliation{%
\institution{Zhejiang University}
\department{School of Software Technology}
\city{Ningbo}
\state{Zhejiang}
\country{China}
}
\email{2840126Z@student.gla.ac.uk}
\author{Yangyang Wu}
\authornote{Corresponding author.}
\affiliation{%
\institution{Zhejiang University}
\department{School of Software Technology}
\city{Ningbo}
\state{Zhejiang}
\country{China}}
\email{zjuwuyy@zju.edu.cn}
\author{Xiaoye Miao}
\affiliation{%
\institution{Zhejiang University}
\department{Center for Data Science}
\city{Hangzhou}
\state{Zhejiang}
\country{China}}
\email{miaoxy@zju.edu.cn}
\author{Mengying Zhu}
\affiliation{%
\institution{Zhejiang University}
\department{School of Software Technology}
\city{Ningbo}
\state{Zhejiang}
\country{China}}
\email{mengyingzhu@zju.edu.cn}
\author{Meng Xi}
\affiliation{%
\institution{Zhejiang University}
\department{School of Software Technology}
\city{Ningbo}
\state{Zhejiang}
\country{China}}
\email{ximeng@zju.edu.cn}
\author{Guanjie Cheng}
\affiliation{%
\institution{Zhejiang University}
\department{School of Software Technology}
\city{Ningbo}
\state{Zhejiang}
\country{China}}
\email{zhaoxinkui@zju.edu.cn}
%%
%% 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}{Trovato et al.}
%%
%% The abstract is a short summary of the work to be presented in the
%% article.
\begin{abstract}
Multi-view multi-label classification (MvMLC) is indispensable for modern web applications aggregating information from diverse sources. However, real-world web-scale settings are rife with missing views and continuously emerging classes, which pose significant obstacles to robust learning.
Prevailing methods are ill-equipped for this reality, as they either lack adaptability to new classes or incur exponential parameter growth when handling all possible missing-view patterns, severely limiting their scalability in web environments.
To systematically address this gap, we formally introduce a novel task, termed \emph{incomplete multi-view multi-label class incremental learning} (IMvMLCIL), which requires models to simultaneously address heterogeneous missing views and dynamic class expansion.
To tackle this task, we propose \textsf{E2PL}, an Effective and Efficient Prompt Learning framework for IMvMLCIL.
\textsf{E2PL} unifies two novel prompt designs: \emph{task-tailored prompts} for class-incremental adaptation and \emph{missing-aware prompts} for the flexible integration of arbitrary view-missing scenarios.
To fundamentally address the exponential parameter explosion inherent in missing-aware prompts, we devise an \emph{efficient prototype tensorization} module, which leverages atomic tensor decomposition to elegantly reduce the prompt parameter complexity from exponential to linear w.r.t. the number of views.
We further incorporate a \emph{dynamic contrastive learning} strategy explicitly model the complex dependencies among diverse missing-view patterns, thus enhancing the model's robustness.
Extensive experiments on three benchmarks demonstrate that \textsf{E2PL} consistently outperforms state-of-the-art methods in both effectiveness and efficiency.
The codes and datasets are available at https://anonymous.4open.science/r/code-for-E2PL.
\end{abstract}
\begin{CCSXML}
<ccs2012>
<concept>
<concept_id>10010147.10010257.10010258.10010259.10010263</concept_id>
<concept_desc>Computing methodologies~Supervised learning by classification</concept_desc>
<concept_significance>500</concept_significance>
</concept>
</ccs2012>
\end{CCSXML}
\ccsdesc[500]{Computing methodologies~Supervised learning by classification}
%%
%% 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]{Do Not Use This Code~Generate the Correct Terms for Your Paper}
% \ccsdesc[300]{Do Not Use This Code~Generate the Correct Terms for Your Paper}
% \ccsdesc{Do Not Use This Code~Generate the Correct Terms for Your Paper}
% \ccsdesc[100]{Do Not Use This Code~Generate the Correct Terms for Your Paper}
% \begin{CCSXML}
% <ccs2012>
% <concept>
% <concept_id>10010147.10010257.10010258.10010259.10010263</concept_id>
% <concept_desc>Computing methodologies~Supervised learning by classification</concept_desc>
% <concept_significance>500</concept_significance>
% </concept>
% </ccs2012>
% \end{CCSXML}
% \ccsdesc[500]{Computing methodologies~Supervised learning by classification}
%%
%% Keywords. The author(s) should pick words that accurately describe
%% the work being presented. Separate the keywords with commas.
\keywords{Incomplete multi-view data, multi-label classification, class incremental learning, prompt learning, contrastive learning}
%% 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
\input{1.Introduction}
\input{2.Related-Work}
\input{3.Method}
\input{4.Experiment}
\input{5.Conclusion}
\balance
\bibliographystyle{ACM-Reference-Format}
\bibliography{main}
\appendix
\input{6.Appendix}
\end{document}
\label{sec:Introduction}
Multi-view multi-label classification (MvMLC) has emerged as a cornerstone technique for web applications that seek to integrate heterogeneous information from multiple sources and capture intricate semantic dependencies among labels~\cite{zhang2018latent, zhao2022non, li2025semi}.
By combining distinct feature representations across views, MvMLC not only enhances predictive accuracy but also increases robustness compared to its single-view counterparts, enabling a wide range of applications such as multimedia retrieval, medical diagnosis, and remote sensing.
In practice, however, multi-view data collected from real-world web environments are often incomplete due to sensor malfunctions, occlusions, user privacy restrictions, or prohibitive acquisition costs.
This leads to the incomplete multi-view multi-label classification (IMvMLC) problem, which has attracted significant attention for its practical importance and inherent complexity~\cite{liu2023dicnet,liu2025reliable, wen2023deep}.
Despite notable progress, existing IMvMLC solutions still encounter two core obstacles that hinder large-scale deployment in dynamic web systems:
\emph{Class-Incremental Flexibility Barrier (\textbf{CH1}):} The prevailing paradigm~\cite{zhao2022learning, liu2023label, liu2024attention} is grounded in static multi-label learning, where the mapping from input features to labels is constructed statically and assumed fixed during deployment as depicted in Figure~\ref{fig:static} (a). However, in open-world web applications, the continual introduction of new categories is the norm rather than the exception.
For instance, in medical image analysis, the discovery of novel disease subtypes is an ongoing process due to advances in research and the accumulation of clinical data \cite{kumar2018example, chen2024classifier}. This necessitates that deployed models not only retain knowledge of previously seen categories, but also incrementally incorporate and recognize new classes as they arise, often without access to the original training data and while avoiding catastrophic forgetting illustrated in Figure~\ref{fig:static} (b). Unfortunately, existing IMvMLC methods, which are typically built on static learning assumptions, lack the flexibility to efficiently accommodate such incremental categories. This incremental inadaptability poses a critical barrier, severely limiting the practical deployment of these models in evolving web environments.
\emph{Scalability and Parameter Explosion (\textbf{CH2}):} Existing IMvMLC methods predominantly rely on complex network architectures, such as autoencoder-based frameworks \cite{liu2023masked, li2025semi, liu2024partial}, which, while effective in handling missing data, often incur high computational overhead and challenging optimization problems. Their rigid designs fundamentally limit scalability and adaptability in scenarios with continuously expanding categories. Recently, prompt-based methods \cite{lee2023multimodal, zheng2024decomposed, lang2025retrieval} have been proposed as lightweight and potentially scalable alternatives for incomplete data learning. However, these methods require designing specific prompt templates for all possible missing-view patterns, resulting in exponential complexity, specifically $\mathcal{O}(2^n)$, where $n$ is the number of views and $2^n$ corresponds to all possible missing-view patterns. This inevitably leads to a parameter explosion.
To address these critical challenges, we introduce a novel task of \emph{incomplete multi-view multi-label class incremental learning} (IMvMLCIL), which demands models to jointly manage heterogeneous missing views and dynamically expanding label spaces. To tackle this, we present \textsf{E2PL}—an Effective and Efficient Prompt Learning framework that unifies both \emph{task-tailored prompts} (TTPs) for swift adaptation to new classes and \emph{missing-aware prompts} (MAPs) for flexible integration of arbitrary missing-view patterns.
% Importantly, unlike existing works, we deliberately retain the expressiveness of assigning a dedicated prompt to each possible missing-view configuration, which maximally preserves adaptation capacity.
To address the associated parameter explosion in MAPs, we introduce an \emph{efficient prototype tensorization} (EPT) module based on atomic tensor decomposition, compressing the prompt space from exponential to linear complexity. Furthermore, the \emph{dynamic contrastive learning} (DCL) strategy is employed to capture nuanced interrelations among diverse missing-view patterns, further enhancing robustness and adaptability.
In summary, our paper makes several notable contributions:
% \begin{itemize}
% \item We formally define the novel task of Incomplete Multi-view Multi-label Class Incremental Learning (IMvMLCIL), which simultaneously addresses the challenges of missing views and the continual emergence of new classes in multi-label settings.
% \item We propose \textsf{E2PL}, an effective and efficient prompt learning framework that introduces both Task-Tailored Prompts for class-incremental adaptation and Missing-Aware Prompts for flexible handling of incomplete data.
% \item We design an Efficient Prototype Tensorization module based on atomic tensor decomposition, which reduces the parameter complexity from exponential $\mathcal{O}(2^n)$ to linear $\mathcal{O}(n)$ with respect to the number of views.
% \item Extensive experiments on three benchmark datasets, demonstrating that \textsf{E2PL} consistently outperforms state-of-the-art methods in both effectiveness and efficiency, providing a practical and scalable solution for real-world IMvMLCIL scenarios.
% \end{itemize}
\begin{itemize}
\item Towards the IMvMLCIL task, we propose a prompt-based framework \textsf{E2PL}, effectively addressing the challenges of heterogeneous missing views and dynamically expanding label spaces.
To the best of our knowledge, this is the first work to holistically address the IMvMLCIL problem within a unified framework.
\item We design an \emph{efficient prototype tensorization} module via atomic tensor decomposition, reducing the parameter complexity of missing-aware prompts from exponential $\mathcal{O}(2^n)$ to linear $\mathcal{O}(n)$ w.r.t.\ the number of views.
\item We propose a \emph{dynamic contrastive learning} strategy to capture nuanced interrelations among diverse missing-view patterns, further enhancing the model’s robustness and adaptability in incremental learning scenarios.
\item Extensive experiments on three benchmark datasets, demonstrating that \textsf{E2PL} consistently outperforms state-of-the-art methods in both effectiveness and efficiency.
\end{itemize}
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{image/f_static_incremental.pdf}
\caption{(a) Illustration of Static Multi-Label Learning, where the model is trained and evaluated on a fixed set of classes. (b) Illustration of Multi-Label Class Incremental Learning, where new classes (e.g., “dog”, “person”, “boat”) are introduced sequentially, requiring the model to adapt to an expanding class set while retaining prior knowledge.}
\label{fig:static}
\vspace{-0.2in}
\end{figure}
\section{Related Work}
\label{sec:Related}
% \subsection{Multi-view Multi-Label Classification}
% Multi-Label classification is a classic and fairly mature task in machine learning. With the advancement of Multi-view learning, increasing attention has been paid to Multi-view Multi-Label Classification (MvMLC). For example, \citet{zhang2018latent} presented a novel MvMLC method based on matrix factorization, where latent patterns across views are aligned in the kernel space. \citet{wu2019multi} proposed SIMM, which combines shared subspace exploitation through the minimization of confusion adversarial loss and multi-label loss with view-specific information extraction via an orthogonal constraint. \citet{zhao2021consistency} developed an innovative method called CDMM, which learns independent prediction results for each view and maximizes the dependence between features and labels through Hilbert–Schmidt Independence Criterion.
% Considering the incompleteness of multi-view data from both feature and label perspectives brings further challenges to MvMLC. Incomplete Multi-view Multi-label Classification (IMvMLC), as proposed by \citet{tan2018incomplete}, first maps incomplete data from multiple views to a shared subspace, and then bridges the shared subspace and the semantic label space using a projection matrix and learnable label correlations. Furthermore, \citet{liu2023dicnet} developed a deep framework named DICNet, which utilizes autoencoders to extract representations and employs contrastive learning on instances to enhance consensus representation ability. The TSIEN model proposed by \citet{tan2024two} aims to maximize task-relevant information while minimizing task-irrelevant information based on information bottleneck theory. \citet{yan2025incomplete} introduced DRLS, which effectively utilizes both consistent and specific information from multi-view data by separating them through a disentangled representation learning module for multi-label classification tasks.
% However, most current research follows the conventional training-testing paradigm of deep learning, where models learn correlations between labels and latent features of multi-view inputs on a fixed training set, and then apply this knowledge to test samples. In real-world scenarios, however, the categorization of objects is continually refined and diversified, leading to an ever-expanding set of label categories. This dynamic environment renders traditional models inadequate for adapting to incremental changes in label sets. To address this limitation, we introduce the concept of \emph{Multi-Label Class Incremental Learning} into the domain of MvMLC.
% \subsection{Multi-View Multi-Label Classification}
% Multi-label classification is a fundamental task in machine learning, and recent advances in multi-view learning have spurred considerable interest in Multi-view Multi-label Classification (MvMLC). For instance, \citet{zhang2018latent} introduced a matrix factorization-based MvMLC approach, aligning latent representations across views in a kernelized space. \citet{wu2019multi} proposed SIMM, which jointly exploits a shared subspace via adversarial and multi-label loss, and captures view-specific features using orthogonal constraints. \citet{zhao2021consistency} presented CDMM, which learns independent predictions for each view and maximizes feature-label dependence through the Hilbert–Schmidt Independence Criterion. The challenge of data incompleteness in both features and labels further complicates MvMLC. Incomplete Multi-view Multi-label Classification (IMvMLC), first studied by \citet{tan2018incomplete}, projects incomplete multi-view data into a common subspace, then bridges it to the label space via a projection matrix and learnable label correlations. \citet{liu2023dicnet} proposed DICNet, a deep framework that combines autoencoder-based representation learning with instance-level contrastive learning for robust consensus representation. The TSIEN model~\cite{tan2024two} leverages information bottleneck theory to disentangle task-relevant and task-irrelevant information for multi-label prediction. Similarly, \citet{yan2025incomplete} developed DRLS, which effectively separates and exploits both consistent and view-specific representations via disentangled learning. Despite these advances, two critical challenges remain insufficiently addressed: Incremental Inadaptability (\textbf{CH1}) and Scalability and Parameter Explosion (\textbf{CH2}). These challenges limit the practical deployment of IMvMLC in dynamic and evolving real-world scenarios.
\subsection{Multi-View Multi-Label Classification}
Multi-view multi-label classification (MvMLC) has emerged as a pivotal research direction, situated at the intersection of multi-view learning and multi-label classification.
Its recent prominence is largely propelled by significant advances in multi-view representation learning.
Zhang et al.~\cite{zhang2018latent} introduced a matrix factorization approach to align latent representations across views.
\citet{wu2019multi} proposed SIMM to jointly exploit a shared subspace and captureview-specific features through adversarial, multi-label lossand orthogonal constraints.
Zhao et al.~\cite{zhao2021consistency} presented CDMM, which learns independent predictions for each view and maximizes feature-label dependence via the Hilbert–Schmidt Independence Criterion. The challenge of data incompleteness in both features and labels adds further complexity.
Incomplete Multi-view Multi-label Classification (IMvMLC), first addressed by Tan et al.~\cite{tan2018incomplete}, projects incomplete data into a common subspace and bridges it to the label space with a projection matrix and learnable label correlations. Subsequent work has improved robustness and representation: DICNet~\cite{liu2023dicnet} employs autoencoders and instance-level contrastive learning; TSIEN~\cite{tan2024two} disentangles task-relevant and irrelevant information using information bottleneck theory; and DRLS~\cite{yan2025incomplete} separates and exploits consistent and view-specific representations via disentangled learning. Nevertheless, \emph{class-incremental flexibility barrier (CH1)} and \emph{scalability and parameter explosion (CH2)} remain insufficiently addressed, hindering the real-world deployment in dynamic environments.
\begin{figure*}[h]
\includegraphics[width=\textwidth]
{image/f1v2.pdf}
\caption{The overall architecture of our method. (a) Training pipeline for Task~$t$: Incomplete multi-view data undergoes linear projection and [CLS] token concatenation. Task-Tailored Prompts and Missing-Aware Prompts are added to the $\mathrm{K}$ and $\mathrm{V}$ vectors, followed by multi-label classification via the classification head. Missing-Aware Prompts are generated using the Efficient Prototype Tensorization module. (b) Inference pipeline: Forward computation for all tasks is performed in parallel, and final predictions are obtained by logit concatenation.}
\label{fig:architecture}
\vspace*{-0.15in}
\end{figure*}
% \subsection{Multi-Label Class Incremental Learning}
% Multi-Label Class Incremental Learning (MLCIL) aims to develop unified models capable of continually integrating new classes while retaining previously acquired knowledge. Recent advances include AGCN and its variant~\cite{du2022agcn, du2023multi}, which leverage augmented Graph Convolutional Networks to model class co-occurrence and incrementally expand graph structures during learning. PRS~\cite{kim2020imbalanced} and OCDM~\cite{liang2022optimizing} enhance online sampling strategies to address class imbalance by maintaining data buffers that better approximate the real distribution. However, as these methods are primarily designed for online scenarios where each batch is encountered only once, their performance degrades in offline settings~\cite{dong2023knowledge}. To address this, KRT~\cite{dong2023knowledge} formulates MLCIL in the offline context, employing an extensible ICA module and token distillation loss to mitigate catastrophic forgetting. APPLE~\cite{song2024overcoming} establishes a comprehensive MLCIL benchmark and proposes a framework that alleviates forgetting through adaptive pseudo-labeling, cluster sampling, and a class attention decoder. More recently, the Multi-lane framework~\cite{de2024less} introduces a prompt-based approach without rehearsal, and RebLL~\cite{du2025rebalancing} addresses the intrinsic positive-negative imbalance in MLCIL. Despite these advancements, existing methods are restricted to single-view settings and fail to capture the complex and complementary relationships among multiple views, limiting their applicability in real-world multi-view scenarios.
% \subsection{Multi-Label Class Incremental Learning}
% Multi-Label Class Incremental Learning (MLCIL) aims to continually construct a unified classifier that can incrementally learn and integrate knowledge, enabling a comprehensive understanding of class objects within images. For example, AGCN and its variant \cite{du2022agcn, du2023multi} employ an augmented Graph Convolutional Network to capture co-occurrence correlations among classes and incrementally expand the graph structure during the learning process. PRS \cite{kim2020imbalanced} and OCDM \cite{liang2022optimizing} enhance sampling methods for online data streams, aiming to sample buffers that better approximate the real data distribution and thus mitigate class imbalance. However, these methods are designed for online learning settings—where each data batch is accessed only once—and tend to exhibit suboptimal performance in offline settings \cite{dong2023knowledge}. To address this, KRT \cite{dong2023knowledge} defines a more promising task, MLCIL, and utilizes an extensible ICA module with token distillation loss to store and transfer knowledge, thereby mitigating catastrophic forgetting. APPLE \cite{song2024overcoming} introduces a novel MLCIL benchmark protocol and presents a framework that addresses catastrophic forgetting through an adaptive pseudo-label strategy, a cluster sampling method, and a class attention decoder. Additionally, \citet{de2024less} proposed the Multi-lane framework, a prompt-based approach that does not require rehearsal. \citet{du2025rebalancing} were the first to identify the inherent positive-negative imbalance in MLCIL and proposed the RebLL framework to address this issue. Nevertheless, existing approaches focus solely on single-view scenarios and are not applicable to multi-view learning tasks. As a result, they fail to capture the complex and deep relationships among multiple views.
% In the single-label CIL. a variety of effective strategies have been proposed to address catastrophic forgetting. Regularization-based methods \cite{kirkpatrick2017overcoming, zenke2017continual} introduced constraints into the loss function to preserve important parameters for old tasks. Replay-based approaches \cite{douillard2020podnet, buzzega2020dark} enhance retention by mixing samples from previous tasks during training, For instance, \citet{riemer2018learning} established a memory buffer to store samples from previous tasks for retraining with new data, while DER++ \cite{buzzega2020dark} suggests applying knowledge distillation penalties to the stored data. Architectural-based methods provide distinct parameters for each task to prevent forgetting. Recent architectural approaches \cite{} combined with the rehearsal methods achieve a better antiforgetting effect. These methods dynamically expand or prune the network parameters to accommodate the new data at the expense of limited scalability.
% \subsection{Prompt Learning}
% Prompt Learning \cite{lester2021power, liang2022modular, khattak2023maple} refers to the process of designing or generating effective prompts to adapt pre-trained models for various downstream tasks. Recently, integrating prompt-tuning techniques into multimodal learning under general missing modality scenarios—where different missing cases are treated as distinct learning tasks—has proven to be an effective and lightweight approach. For instance, \citet{lee2023multimodal} proposed multimodal prompting, which utilizes learnable tokens called missing-aware prompts (MAPs) to transfer knowledge from pretrained models on large-scale datasets to downstream tasks. Similarly, \citet{guo2024multimodal} introduced a prompt learning framework for sentiment analysis and emotion recognition that is both computationally efficient and capable of handling missing modalities during both training and testing.
% However, these approaches require designing separate prompts for each missing pattern, causing the number of parameters to grow exponentially with the number of modalities. To address this, \citet{jang2024towards} proposed Modality-Specific Prompts (MSPs), which assign independent prompts to each modality, thereby reducing parameter complexity from exponential to linear. Nevertheless, as the number of modalities increases, this straightforward design struggles to capture the complex combinations of missing modalities. In this work, we retain the strategy of designing independent prompts for each missing case, but further decompose and reduce parameter complexity from a novel perspective, achieving optimal performance with minimal parameter cost.
\subsection{Prompt Learning}
Prompt learning~\cite{lester2021power, liang2022modular, khattak2023maple} adapts pre-trained models to diverse downstream tasks by designing effective prompts. Recently, the integration of prompt learning techniques into multimodal learning under the general missing modality scenario---where each specific missing pattern is regarded as a distinct task---has emerged as a lightweight and efficient adaptation strategy \cite{lee2023multimodal, guo2024multimodal, lang2025retrieval}. Despite significant progress, existing approaches typically require separate prompts for each missing modality pattern, leading to exponential growth in parameters as the number of modalities increases. \citet{jang2024towards} mitigate this by introducing Modality-Specific Prompts, reducing parameter complexity from exponential to linear by assigning independent prompts to each modality. Building on this, \citet{chen2024epe} proposed EPE-P, which further improves efficiency and performance by integrating cross-modal prompts.
However, as the number of modalities increases, such independent modeling approaches fail to capture the nuanced dependencies across missing scenarios, ultimately limiting their adaptability and expressiveness.
In this work, we retain the independent prompt design for each missing case and further propose a parameter decomposition technique that minimizes parameter overhead while achieving optimal performance.
\section{Methodology}
\label{sec:Overview}
\subsection{Problem Definition}
Assume there are a total of $T$ incremental sessions $\{ (\mathcal{X}^1, \mathcal{Y}^1, \mathcal{C}^1), \cdots \\ (\mathcal{X}^T, \mathcal{Y}^T, \mathcal{C}^T) \}$, where each session $t$ provides a training set $(\mathcal{X}^t, \mathcal{Y}^t)$ and a newly introduced, disjoint label set $\mathcal{C}^t$ (i.e., $\mathcal{C}^m \cap \mathcal{C}^n = \emptyset$ for $m \neq n$). For each session, $\mathcal{X}^t = \left\{ \left( x_{i,1}^t, \cdots, x_{i,n}^t \right), m_i^t \right\}_{i=1}^{N_t}$ denotes the multi-view features of $N_t$ samples, where $n$ is the number of views, $x_{i,v}^t$ is the feature vector of the $v$-th view, and $m_i^t \in \{0,1\}^n$ is a view-indicator vector with $m_{i,v}^t=1$ if the $v$-th view is available and $0$ otherwise, allowing for missing views. The corresponding label set $\mathcal{Y}^t = \{ y_i^t \}_{i=1}^{N_t}$ contains multi-label annotations for the current session, where each $y_i^t \in \{0,1\}^{|\mathcal{C}^t|}$.
The goal of \emph{incomplete multi-view multi-label class incremental learning} (IMvMLCIL) is to incrementally train a unified multi-label classification model that can accurately predict labels over all encountered classes for samples with missing multi-view features.
At each session $t$, only the current training data $(\mathcal{X}^t, \mathcal{Y}^t)$ are available for model update. After session $t$, the model is evaluated on a test set covering all classes seen so far, i.e., the cumulative class space $\mathcal{C}^{1:t} = \bigcup_{\tau=1}^{t} \mathcal{C}^\tau$.
For each test sample, the model must predict a multi-label vector over $\mathcal{C}^{1:t}$ based on its possibly incomplete multi-view features. During preprocessing, missing values in $\mathcal{X}$ are imputed with zeros.
\subsection{Prompt Design}
Existing IMvMLC methods \cite{liu2023masked, li2025semi, liu2024partial} mainly use autoencoder-based models to extract unimodal and multimodal features, which are fused and classified by multi-layer perceptrons.
While effective, these architectures are often cumbersome and lack scalability and adaptability, especially in class-incremental scenarios common to real-world applications. To address these limitations, we propose a unified prompt-based framework that leverages the efficiency and flexibility of prompt learning, as illustrated in Figure~\ref{fig:architecture}.
Our proposed framework \textsf{E2PL} orchestrates two complementary prompt types—task-tailored for class increments and missing-aware for data incompleteness—to achieve robust performance in dynamic and imperfect data environments.
\emph{Task-Tailored Prompts} (TTPs).
To support class incremental learning scenario, we structure the problem as a sequence of tasks, each bringing in new classes.
For each task, we design TTPs specific to its requirements.
Formally, let $P_T = \{P_T^1, P_T^2, \cdots, P_T^T\}$, where $P_T^t$ denotes the prompt uniquely learned for the $t$-th task. These prompts are attached to a frozen, pre-trained transformer backbone, allowing efficient injection of task-specific knowledge without retraining the model. During training for the $t$-th task, only $P_T^t$ is updated, while prompts from previous tasks remain fixed to prevent catastrophic forgetting illustrated in Figure~\ref{fig:architecture} (a).
At inference (Figure~\ref{fig:architecture} (b)), all task-tailored prompts are processed in parallel, and the outputs from the corresponding task-specific classification heads are concatenated for final prediction.
% This design enables the model to adapt efficiently and flexibly to new tasks via modular and lightweight prompts.
\emph{Missing-Aware Prompts} (MAPs).
To address the challenge posed by varying missing-view configurations in incomplete multi-view data, we propose Missing-Aware Prompts (MAPs), where each prompt is explicitly engineered to represent a unique pattern of view availability. For $n$ input views, there exist $2^n-1$ possible missing-view patterns, as each view can be either present or absent (for convenience, we use $2^n$ in the following discussion). Ideally, each pattern would be assigned a unique prompt, denoted as $P_M = \{P_M^1, \cdots, P_M^{2^n}\}$, where $P_M^i$ corresponds to the $i$-th missing-view pattern.
% However, this straightforward approach leads to an exponential parameter complexity of $\mathcal{O}(2^n)$, which is impractical for large $n$. To surmount this scalability bottleneck, we introduce a prototype tensorization module that dynamically constructs MAPs from a compact, learnable set of prototypes. The technical details of this module are presented in the next section.
% To handle incomplete multi-view scenarios, we introduce MAPs that explicitly encode each missing-view pattern.
% Given $n$ input views, there are $2^n - 1$ possible non-trivial missing-view patterns (for convenience, we use $2^n$ in the following discussion).
% Ideally, each missing pattern would have its unique prompt, forming $P_M = \{P_M^1, \cdots, P_M^{2^n}\}$, where each $P_M^i$ corresponds to a specific missing-view pattern.
% {\color{red}However, this naive solution incurs an exponential parameter complexity of $\mathcal{O}(2^n)$, which quickly becomes infeasible for large $n$.
% To address this, we propose an \emph{efficient prototype tensorization} module, which compresses the prompt space and reduces complexity to $\mathcal{O}(n)$ by dynamically generating the appropriate missing-aware prompt according to the current missing-view mask.
% The technical details of this module are presented in the next section.}
% \subsection{Prompt Design}
% Existing methods for incomplete multi-view multi-label learning mainly rely on autoencoder-based models to extract unimodal and multimodal representations, which are then fused and classified using multi-layer perceptrons. While effective, such approaches often result in complex and cumbersome architectures, limiting their scalability and adaptability to the increasingly common class-incremental scenarios encountered in real-world applications. To address these challenges, we propose a unified prompt-based framework that leverages the efficiency and scalability of prompt learning, as illustrated in Figure \ref{fig:architecture}. Our approach introduces two complementary prompt modules: Task-Tailored Prompts and Missing-Aware Prompts, which together enable the model to flexibly handle both class-incremental and incomplete multi-view challenges.\\
% \textbf{Task-Tailored Prompts.}
% To support class-incremental learning, we formulate the problem as a sequence of tasks, each introducing new classes. For each task, we design a tailored prompt---that is, a prompt specifically customized for the characteristics and requirements of that particular task. Formally, we denote the set of task-tailored prompts as $P_T = \{P_T^1, P_T^2, \cdots, P_T^T\}$, where each $P_T^t$ is uniquely learned for the $t$-th task. These tailored prompts are attached to a frozen, pre-trained transformer backbone, enabling the model to efficiently adapt to new classes by injecting task-specific knowledge, without the need to retrain the entire network. During the $t$-th task session, only the current tailored prompt $P_T^t$ is updated, while prompts for previous tasks remain frozen to preserve knowledge from earlier sessions. As shown in Figure~1(a), each Task-Tailored Prompt is injected into the transformer, and during inference (Figure~1(b)), all task-tailored prompts are processed in parallel. The final prediction is obtained by concatenating the outputs from the task-specific classification head. In this way, our approach ensures that the model can flexibly and efficiently adapt to each new task through a set of tailored, lightweight prompts.\\
% \textbf{Missing-Aware Prompts.}
% To address the challenge of incomplete multi-view data, we introduce Missing-Aware Prompts that explicitly encode the patterns of missing views. Given $n$ input views, there are $2^n - 1$ possible non-empty missing-view configurations ( For convenience, we utilize $2^n$ in the following sections.). Ideally, we would assign a unique prompt to each missing-view pattern, allowing the model to perceive and adapt to any scenario of missing data. That is, the set of Missing-Aware Prompts can be represented as $P_M = \{P_M^1, P_M^2, \cdots, P_M^{2^n}\}$, where each $P_M^i$ corresponds to a specific missing-view configuration. However, this straightforward approach leads to an exponential space complexity of $\mathcal{O}(2^n)$, which is impractical for large $n$. To overcome this limitation, we propose the \textbf{Efficient Prototype Tensorization} module, which compresses the prompt space and reduces the complexity to $\mathcal{O}(n)$ by dynamically generating the appropriate Missing-Aware Prompt based on the current missing-view mask. The details of this module will be introduced in the next section.
\vspace*{-0.1in}
\subsection{Efficient Prototype Tensorization}
% {\color{red}Prompt Learning has demonstrated strong adaptability for incomplete multi-modal or multi-view data. However, as the number of modalities or views increases, the required prompts grow exponentially, leading to significant parameter overhead.
% To address this limitation, we propose an \emph{efficient prototype tensorization} (EPT) module.}
% Our approach preserves the expressive power of generating a distinct prompt for each missing-view pattern, while leveraging atomic tensor decomposition to reduce parameter complexity from exponential to linear in $n$.
% This enables the model to effectively encode intricate dependencies across missing-view patterns. Given $n$ input views, there are $2^n$ possible missing-view patterns, as each view may be present or absent. To explicitly model these patterns, we define the Missing-Aware Prompts matrix $P_M \in \mathbb{R}^{d \times 2^n}$, where $d$ is the prompt length and each column $P_M^i \in \mathbb{R}^d$ corresponds to the missing-aware prompt for a specific missing-view pattern, $i \in [2^n]$. While this formulation is fully expressive, the parameter complexity $\mathcal{O}(d \cdot 2^n)$ becomes prohibitive as $n$ grows. To overcome this, we adopt a structured parameterization that combines low-rank matrix factorization \cite{sainath2013low} with Tensor Train (TT) decomposition \cite{oseledets2011tensor}.
% {\color{blue}Prompt learning has demonstrated strong adaptability for incomplete multi-view data, but assigning a unique prompt to each of the $2^n$ missing-view patterns results in exponential parameter complexity.
% \textcolor{blue}{
% In MAPs, each missing-view configuration is assigned a distinct prompt, resulting in $2^n$ possible patterns and an exponential parameter complexity of $\mathcal{O}(2^n)$ for $n$ input views. This rapid parameter escalation renders MAPs impractical for large-scale scenarios, fundamentally constraining the scalability in real-world multi-view applications.}
The direct parameterization of MAPs presents a fundamental scalability bottleneck. By assigning a distinct prompt to each of the $2^n$ possible missing-view configurations for $n$ input views, the approach incurs an exponential parameter complexity of $\mathcal{O}(2^n)$, rendering it intractable for real-world applications with a large number of views. To mitigate the parameter explosion, we propose an \emph{efficient prototype tensorization} (EPT) module, which leverages atomic tensor decomposition to reduce the parameter complexity from exponential to linear in $n$, i.e., $\mathcal{O}(n)$, while preserving the expressiveness of distinct prompts.
Specifically, we define a MAPs matrix $P_M \in \mathbb{R}^{d \times 2^n}$, where $d$ is the prompt length and each column $P_M^i$ represents the prompt for the $i$-th missing-view pattern. To ensure tractability, we employ a \emph{structured parameterization} that integrates low-rank matrix factorization~\cite{sainath2013low} and \emph{tensor train} (TT) decomposition~\cite{oseledets2011tensor}, enabling efficient modeling of complex dependencies among missing-view patterns.
Specifically, we reparameterize the MAPs matrix using a factorized structure:
\begin{equation}
P_M = A \, B^\top,
\end{equation}
where $A \in \mathbb{R}^{d \times k}$ is a shared basis matrix ($k \ll d$), and $B \in \mathbb{R}^{2^n \times k}$ contains the coefficients for each missing-view pattern. This reduces the total number of parameters from $\mathcal{O}(d \cdot 2^n)$ to $\mathcal{O}((d + 2^n) \cdot k)$.
However, the size of $B$ still scales exponentially in $n$. To further compress $B$, we reshape it into a tensor $\mathcal{B} \in \mathbb{R}^{2 \times 2 \times \cdots \times 2 \times k}$ of order $n+1$, where each of the $n$ binary modes corresponds to the presence ($1$) or absence ($0$) of a view and the last mode indexes the k latent factors. We then apply TT decomposition to $\mathcal{B}$ which expresses it as a sequence of core tensors $\{G_1, \cdots , G_{n+1}\}$, where $G_\ell \in \mathbb{R}^{r_{\ell-1} \times 2 \times r_\ell}$ for $\ell<=n$ and $G_{n+1} \in \mathbb{R}^{r_n \times k}$. The TT decomposition is defined as:
\begin{equation}
\mathcal{B} (s_1, \cdots, s_n) = G_1(s_1)\cdots G_n(s_n) \cdot G_{n+1},
\end{equation}
where $s_j \in \{0, 1\}$ indicates the presence of the $j$-th view, and $G_j(s_j) \in \mathbb{R}^{r_{j-1} \times r_j}$ denotes the $s_j$-th slice of the $j$-th core tensor. The resulting coefficient vector for a missing-view pattern $m = (m_1, \cdots, m_n)$ is:
\begin{equation}
\beta_m = G_1(m_1) \cdots G_n(m_n) \cdot G_{n+1}.
\end{equation}
Finally, the MAP for pattern $m$ is obtained as:
\begin{equation}
P_M^m = A \, \beta_m^\top.
\end{equation}
This structured parameterization enables EPT to efficiently model complex dependencies among missing-view patterns while maintaining scalability. The total number of learnable parameters becomes $\mathcal{O}(n \cdot R^2 \cdot k + d \cdot k)$, where $R = \max(r_0, \cdots, r_n)$ is the maximum TT rank. This reduces the complexity from exponential to linear in the number of views $n$ , making EPT suitable for high-dimensional multi-view settings.
% Moreover, the TT structure enables implicit knowledge sharing among related missing-view patterns, thereby improving generalization to unseen missing-view cases.
% \subsection{Efficient Prototype Tensorization}
% Prompt learning has demonstrated strong effectiveness in handling incomplete multi-modal (or multi-view) data. However, as the number of modalities (or views) increases, the number of required prompts grows exponentially, resulting in significant parameter complexity. Existing approaches typically address this by constructing a dedicated prompt for each modality and then combining these prompts to adapt to all possible missing-view configurations. While this strategy reduces the parameter count, simple linear combinations of prompts are insufficient to capture the complex dependencies among missing modalities as the number of views $n$ increases.
% To overcome this limitation, we propose the Efficient Prototype Tensorization module. Our approach retains the idea of generating a unique prompt for each missing-view configuration, but leverages atomic tensor decomposition to reduce parameter complexity from exponential to linear in $n$. This enables the model to efficiently capture intricate dependencies among different missing-view configurations.
% Given $n$ input views, there exist $2^n$ distinct missing-view configurations, as each view can be either present or absent. To explicitly model the unique characteristics of each configuration, we construct a Missing-Aware Prompts matrix $P_M \in \mathbb{R}^{d \times 2^n}$, where $d$ denotes the prompt length. Each column $P_M^i \in \mathbb{R}^d$ corresponds to the missing-aware prompt for the $i$-th missing-view configuration, where $i \in [2^n]$. While this formulation ensures comprehensive coverage of all possible missing-view configurations, it introduces a scalability challenge: the parameter complexity of $P_M$ is $\mathcal{O}(d \cdot 2^n)$, which grows exponentially with the number of views $n$. As $n$ increases, this exponential growth quickly becomes intractable, severely limiting the practicality of this approach in large-scale multi-view learning.
% To address the exponential growth of the Missing-Aware Prompts space, we propose a structured parameterization that combines low-rank matrix decomposition with Tensor Train (TT) decomposition. This enables us to generate missing-aware prompts for any missing-view configuration, reducing the overall parameter complexity from $\mathcal{O}(d \cdot 2^n)$ to $\mathcal{O}(n \cdot R^2 \cdot k + d \cdot k)$, where $R$ is the maximum TT rank and $k$ is the low-rank dimension.
% Specifically, we reparameterize the Missing-Aware Prompts matrix using a factorized structure:
% \begin{equation}
% P_M = A \, B^\top,
% \end{equation}
% where $A \in \mathbb{R}^{d \times k}$ is a shared basis matrix ($k \ll d$), and $B \in \mathbb{R}^{2^n \times k}$ contains the coefficients for each missing-view configuration. This reduces the total number of parameters from $\mathcal{O}(d \cdot 2^n)$ to $\mathcal{O}((d + 2^n) \cdot k)$.
% However, the size of $B$ still scales exponentially with $n$. To further reduce complexity, we apply Tensor Train (TT) decomposition to $B$, treating it as a high-order tensor $\mathcal{B} \in \mathbb{R}^{2 \times 2 \times \cdots \times 2 \times k}$, where each of the $n$ binary modes corresponds to the presence or absence of a view. The TT decomposition expresses $\mathcal{B}$ as a sequence of core tensors $G_1, G_2\cdots, G_n, G_{n+1}$:
% \begin{equation}
% B (s_1, \cdots, s_n) = G_1(s_1)\cdot G_1(s_2)\cdots G_n(s_n) \cdot G_{n+1},
% \end{equation}
% where $s_j \in \{0, 1\}$ indicates whether the $j$-th view is present ($1$) or missing ($0$), $G_\ell(i_\ell) \in \mathbb{R}^{r_{\ell-1} \times r_\ell}$ is the $\ell$-th core tensor at state $i_\ell$, and $G_{n+1} \in \mathbb{R}^{r_n \times k}$ projects the final interaction to the coefficient space. For any given missing-view configuration $m = (m_1, \cdots, m_n)$, the corresponding coefficient vector is computed as:
% \begin{equation}
% \beta_m = G_1(m_1) \cdot G_1(m_2) \cdots G_n(m_n) \cdot G_{n+1}.
% \end{equation}
% The missing-aware prompt for the $i$-th missing-view configuration is then synthesized as:
% \begin{equation}
% P_M^m = A \, \beta_m^\top.
% \end{equation}
% This approach enables efficient and scalable parameterization of all possible Missing-Aware Prompts. The total number of learnable parameters is now $\mathcal{O}(n \cdot R^2 \cdot k + d \cdot k)$, where $R = \max(r_0, \cdots, r_n)$ is the maximum TT rank. Moreover, the TT structure enables implicit knowledge sharing among related missing-view configurations, thereby improving generalization to unseen missing-view configurations.
\subsection{Dynamic Contrastive Learning}
% The relationships among different missing-view patterns in multi-view data are inherently intricate, as the presence or absence of specific views fundamentally shapes the semantic proximity between patterns. Intuitively, missing patterns sharing common observed views should exhibit higher similarity in the latent space. However, a naive modeling approach that solely considers the number of shared views fails to capture the nuanced reality that the contribution of each view to the overall similarity is not uniform—certain views may carry more discriminative or complementary information than others, depending on the context.
The semantic relationships among missing-view patterns in multi-view data are inherently complex, as the presence or absence of specific views critically determines their latent similarity. While patterns sharing observed views are intuitively closer in the latent space, simply counting shared views overlooks the fact that each view contributes unequally—some offer more discriminative or complementary information depending on context.
To address these challenges, we propose a \emph{dynamic contrastive learning} (DCL) strategy that explicitly models both the structural overlap and the heterogeneous importance of each view. Specifically, given $n$ input views, and each possible missing pattern can be represented as a binary vector of length $n$, resulting in $2^n$ possible patterns. We assign a learnable weight parameter $w = [w_1, w_2, ..., w_n]$ to each view. For any two missing patterns $i$ and $j$, represented by binary view-indicator vector $m_i$ and $m_j$, we compute their weighted overlap score as follows:
\begin{equation}
s_{ij} = m_i^\top \cdot \mathrm{Sigmoid}(w) \cdot m_j
\end{equation}
where the $\mathrm{Sigmoid}$ function constrains each weight to $(0, 1)$. Based on $s_{ij}$, we construct positive pairs ($s_{ij} > 0$) that share at least one observed view, and negative pairs ($s_{ij} = 0$) with no overlap. The dynamic contrastive learning loss is designed to pull positive pairs closer in the embedding space while enforcing a margin $\alpha$ between negative pairs:
\begin{align}
\mathcal{L}_{DCL} =\ & \frac{1}{|P|} \sum_{(i,j) \in P} \|P_M^i - P_M^j\|^2 \notag \\
& + \frac{1}{|N|} \sum_{(i,j) \in N} \max\left(0,\, \alpha - \|P_M^i - P_M^j\|\right)^2,
\end{align}
where $P$ and $N$ denote the sets of positive and negative pairs, $\alpha$ is the distance margin, and $\|\cdot\|$ denotes the Euclidean norm.
% ========== 训练算法 ==========
\begin{algorithm}[t]
\caption{Training process of \emph{E2PL}}
\label{alg:E2PL_training}
\begin{algorithmic}[1]
\REQUIRE Incomplete multi-view data $(\mathcal{X}^t, \mathcal{Y}^t)$ for task $t$; view-indicator matrix $M^t$; hyperparameters $\lambda$.
\ENSURE Updated model parameters.
\FOR{each sample $i$ in $\mathcal{X}^t$}
\STATE Impute missing views with zeros according to $m_i^t$
\STATE Extract view-specific features $h_{i,v}^t$ for all views $v$ \hfill (Eq.7)
\STATE Form input sequence $S_i^t = [h_{[\text{CLS}]}, h_{i,1}^t, ..., h_{i,n}^t]$ \hfill (Eq.8)
\STATE Generate MAPs $P_M$ using EPT module
\STATE Index the corresponding MAP $P_M^i$ with $m_i^t$
\hfill (Eq.9)
\STATE Retrieve TTP $P_T^t$ for current task
\STATE Split $P_M^i$ into $P_M^{i,K}, P_M^{i,V}$; split $P_T^t$ into $P_T^{t,K}, P_T^{t,V}$
\hfill (Eq.10)
\STATE Inject $P_M^{i,K}, P_M^{i,V}, P_T^{t,K}, P_T^{t,V}$ into K and V vectors
\hfill (Eq.11)
\STATE Forward pass to obtain representation $z_i^t$ and prediction $\hat{y}_i^t$ \hfill (Eq.12)
\ENDFOR
\STATE Compute binary cross-entropy loss $\mathcal{L}_{BCE}$ \hfill (Eq.13)
\STATE Compute dynamic contrastive learning loss $\mathcal{L}_{DCL}$ \hfill (Eq.6)
\STATE Compute total loss $\mathcal{L}_{Total} = \mathcal{L}_{BCE} + \lambda \cdot \mathcal{L}_{DCL}$ \hfill (Eq.14)
\STATE Update model parameters for current task $t$ (freeze prompts from previous tasks)
\end{algorithmic}
\end{algorithm}
\subsection{Training and Inference Pipeline}
To facilitate reproducibility and provide a clear overview of our approach, we present the complete training and inference workflows in Algorithm~\ref{alg:E2PL_training} and Algorithm~\ref{alg:E2PL_inference}, respectively. These algorithms encapsulate the core mechanisms of E2PL, including data preprocessing, prompt generation and injection, and model update, and are designed to efficiently address both class-incremental adaptation and arbitrary missing-view scenarios.
\emph{Training Pipeline.} Taking task $t$ session as an example, the training pipeline begins with the current session’s dataset $(\mathcal{X}^t, \mathcal{Y}^t)$ and the corresponding label set $\mathcal{C}^t$.
For each sample $i$, the multi-view features $(x_{i,1}^t, \cdots, x_{i,n}^t)$ are first preprocessed by imputing missing views as zero vectors, according to the view-indicator vector $m_i^t \in \{0,1\}^n$. Each available view is then projected into a shared embedding space via a view-specific linear projection encoder:
\begin{equation}
h_{i,v}^t = \mathrm{Encoder}_v(x_{i,v}^t), \quad v=1,\cdots,n.
\end{equation}
The input sequence to the task $t$ forward module, which is composed of $N$ stacked Transformer layers, is formed by concatenating the [CLS] token embedding with the embeddings of all views, yielding
\begin{equation}
S_i^t = [h_{\mathrm{[CLS]}}, h_{i,1}^t, \cdots, h_{i,n}^t].
\end{equation}
To explicitly encode the missing-view pattern, we first use the EPT module to generate MAPs $P_M = \{P_M^1, \cdots, P_M^{2^n}\}$.
For each sample $i$, the view-indicator vector $m_i^t$ is used to index the corresponding missing-aware prompt $P_M^i$ from $P_M$:
\begin{equation}
P_M^i = P_M[m_i^t].
\end{equation}
Meanwhile, for the current task $t$, the corresponding task-tailored prompt $P_T^t$ is retrieved. Both $P_M^i$ and $P_T^t$ are split equally along the feature dimension into key and value parts:
\begin{equation}
P_M^{i,K}, P_M^{i,V} = \mathrm{Split}(P_M^i), \quad
P_T^{t,K}, P_T^{t,V} = \mathrm{Split}(P_T^t)
\end{equation}
At each transformer layer $l$, $P_M^{i,K} $ and $P_M^{i,V}$ are added to the key and value representations, respectively:
\begin{equation}
K^{(l)} = K^{(l)}_{\text{orig}} + P_M^{i,K} + P_T^{t,K}, \quad
V^{(l)} = V^{(l)}_{\text{orig}} + P_M^{i,V} + P_T^{t,V}
\end{equation}
where $K^{(l)}_{\text{orig}}$ and $V^{(l)}_{\text{orig}}$ denote the original key and value matrices computed via $\mathrm{QKV}$ projection, respectively.
The prompt-augmented sequence is then processed through multi-head attention and MLP layers.
Finally, the [CLS] token output from the last layer, denoted as $z_i^t$ used as the global representation of the multi-view sample. This representation is then passed to a task-specific classification head to predict the current class space $\mathcal{C}^{t}$:
\begin{equation}
\hat{y}_i^t = \sigma(W_c^t z_i^t + b_c^t),
\end{equation}
where $\sigma(\cdot)$ is the sigmoid activation function, and $W_c^t$, $b_c^t$ are the parameters of the classification head.
% {\color{red}To guide the model towards accurate classification, we employ the \emph{binary cross-entropy} (BCE) loss over the current task classes:
% \begin{equation*}
% \mathcal{L}_{BCE} = -\frac{1}{|\mathcal{Y}^t|} \sum_{i \in \mathcal{Y}^t} \left[ y_i^t \log \hat{y}_i^t + (1 - y_i^t) \log (1 - \hat{y}_i^t) \right],
% \end{equation*}
% where $y_i^t$ is the ground truth binary label and $\hat{y}_i^t$ is the corresponding model prediction.}
% To further enhance the model's ability to capture the nuanced relationships among different missing-view patterns, we propose a Dynamic Contrastive Learning (DCL) strategy to dynamically model the importance of each view under multi-view missing patterns. Specifically, given $n$ input views, and each possible missing pattern can be represented as a binary vector of length $n$, resulting in $2^n$ possible patterns. We assign a learnable weight parameter $w = [w_1, w_2, ..., w_n]$ to each view. For any two missing patterns $i$ and $j$, represented by binary view-indicator vector $m_i$ and $m_j$, we compute their weighted overlap score as follows:
% \begin{equation}
% s_{ij} = m_i^\top \cdot \mathrm{sigmoid}(w) \cdot m_j
% \end{equation}
% where $\mathrm{sigmoid}$ ensures each view’s weight lies in $(0, 1)$, enabling adaptive importance modeling. Based on $s_{ij}$, we define positive pairs ($s_{ij} > 0$, sharing at least one observed view) and negative pairs ($s_{ij} = 0$, no overlap). Positive pairs are encouraged to have similar embeddings, while negative pairs are separated by a margin $\alpha$. The contrastive loss for Missing-Aware Prompts $P_M$ is:
% % Here, $\mathrm{sigmoid}$ ensures that each view's weight is in the range $(0,1)$, allowing the model to adaptively assess the importance of different views. Based on the weighted overlap score $s_{ij}$, we flexibly define positive and negative sample pairs: if $s_{ij} > 0$, the two missing patterns share at least one observed view and are considered a positive pair, encouraging their embeddings to be close; otherwise, if $s_{ij} = 0$, they have no overlap in observed views and are considered a negative pair, encouraging their embeddings to be farther apart than a margin $\alpha$. We compute the following contrastive loss for the missing-aware prompt $P_M$:
% \begin{align*}
% \mathcal{L}_{cl} =\ & \frac{1}{|P|} \sum_{(i,j) \in P} \|P_M^i - P_M^j\|^2 \notag \\
% & + \frac{1}{|N|} \sum_{(i,j) \in N} \max\left(0,\, \alpha - \|P_M^i - P_M^j\|\right)^2,
% \end{align*}
% where $P$ and $N$ denote the sets of positive and negative pairs, $\alpha$ is the distance margin, and $\|\cdot\|$ denotes the Euclidean norm. The overall training objective is $\mathcal{L}_{\mathrm{bce}}^t + \lambda*\mathcal{L}_\mathrm{cl}$, where $\lambda$ is a hyperparameter that balances two loss terms.
% ========== 推理算法 ==========
\begin{algorithm}[t]
\caption{Inference process of \emph{E2PL}}
\label{alg:E2PL_inference}
\begin{algorithmic}[1]
\REQUIRE Test sample $\mathcal{X}_i$; view-indicator vector $m_i$.
\ENSURE Final prediction $\hat{Y}_i$ over all classes.
\STATE Impute missing views with zeros according to $m_i$
\STATE Extract view-specific features and form input sequence $S_i$
\STATE Generate MAP $P_M^i$ using EPT module with $m_i$
\FOR{each task $\tau = 1$ to $T$}
\STATE Load task-tailored prompt $P_T^\tau$
\STATE Split $P_M^i$ into $P_M^{i,K}$ and $P_M^{i,V}$; split $P_T^\tau$ into $P_T^{\tau,K}$ and $P_T^{\tau,V}$
\STATE Inject $P_M^{i,K}, P_M^{i,V}, P_T^{\tau,K}, P_T^{\tau,V}$ into K and V vectors
\STATE Forward pass to get task-specific logit $\hat{y}_i^\tau$
\ENDFOR
\STATE Concatenate all logits: $\hat{Y}_i = [\hat{y}_i^1; \hat{y}_i^2; ...; \hat{y}_i^T]$ \hfill (Eq.15)
\end{algorithmic}
\end{algorithm}
\emph{Objective Function.}
To simultaneously enhance predictive robustness and representational capacity, we formulate a unified objective function that synergistically combines two distinct loss terms. The first is the standard binary cross-entropy (BCE) loss, which governs performance on the primary classification task, i.e.,
\begin{equation}
\mathcal{L}_{BCE} = -\frac{1}{|\mathcal{Y}^t|} \sum_{i \in \mathcal{Y}^t} \left[ y_i^t \log \hat{y}_i^t + (1 - y_i^t) \log (1 - \hat{y}_i^t) \right],
\end{equation}
where $y_i^t$ is the ground truth binary label and $\hat{y}_i^t$ is the corresponding model prediction.
The second, a dynamic contrastive learning loss $\mathcal{L}_{DCL}$, regularizes the latent space by enforcing semantic consistency and discriminability among diverse missing-view patterns.
Hence, the final objective for \textsf{E2PL}, denoted $\mathcal{L}_{Total}$, is the weighted aggregation of these two components:
\begin{equation}
\mathcal{L}_{Total} =
\mathcal{L}_{BCE} + \lambda \cdot \mathcal{L}_{DCL},
\end{equation}
where $\lambda$ is a hyperparameter that balances two loss terms.
\emph{Inference Pipeline.}
During the inference stage, for a test sample $\mathcal{X}_i$, the MAP $P_M^i$ is generated by the EPT module according to the view-indicator vector $m_i$.
Meanwhile, all TTPs from the $T$ tasks are attached in parallel to the backbone network, resulting in $T$ forward propagation pathways, as shown in Figure~\ref{fig:architecture} (b).
The [CLS] representation from each pathway is fed into its corresponding classification head to obtain the task-specific logit. All logits are then concatenated to form the final prediction vector:
\begin{equation}
\hat{Y} = [\hat{y}^1; \cdots; \hat{y}^T],
\end{equation}
where $[\cdot;\cdot]$ denotes the concatenation operation. The final multi-label prediction is obtained by independently thresholding each sigmoid-normalized logit, which is the standard practice in multi-label and binary classification tasks.
\section{Experiment}
\begin{table*}[ht]
\centering
\setlength{\tabcolsep}{4.5pt}
\caption{Evaluation on three datasets at $R_V = 30\%$. ``Memory'' denotes the number of replay samples; ``-R'' indicates the memory allocated to the method. \textbf{Bold} and \underline{underlined} values indicate the \textbf{best} and \underline{second-best} results, respectively.}
\begin{tabular}{l c c cccc cccc cccc}
\toprule
\multirow{3}{*}{\textbf{Methods}} & \multirow{3}{*}{\textbf{Types}} & \multirow{3}{*}{\textbf{Memory}}
& \multicolumn{4}{c}{\emph{ESPGame}}
& \multicolumn{4}{c}{\emph{IAPRTC12}}
& \multicolumn{4}{c}{\emph{MIRFLICKR}}\\
\cmidrule(lr){4-7} \cmidrule(lr){8-11} \cmidrule(lr){12-15}
& & &\multicolumn{1}{c}{Avg.} & \multicolumn{3}{c}{Last}
&\multicolumn{1}{c}{Avg.} & \multicolumn{3}{c}{Last}
&\multicolumn{1}{c}{Avg.} & \multicolumn{3}{c}{Last}\\
\cmidrule(lr){4-4} \cmidrule(lr){5-7} \cmidrule(lr){8-8} \cmidrule(lr){9-11} \cmidrule(lr){12-12} \cmidrule(lr){13-15}
& & & mAP & CF1 & OF1 & mAP & mAP & CF1 & OF1 & mAP & mAP & CF1 & OF1 & mAP \\
\midrule
Upper-bound & Baseline & 0 & - & 2.99 & 13.54 & 30.17 & - & 3.24 & 17.49 & 32.45 & - & 36.71 & 51.25 & 62.47 \\
\midrule
DICNet & IMvMLC & \multirow{8}{*}{0} & 20.01 & 0.61 & 5.82 & 17.82 & 23.42 & 0.58 & 12.67 & 22.50 & 61.29 & 19.04 & 43.13 & 49.68 \\
MTD & IMvMLC & & 22.89 & 0.95 & 8.92 & 19.79 & 26.98 & 1.23 & 13.30 & 25.12 & 63.00 & 22.34 & 43.66 & 52.11 \\
AIMNet &IMvMLC& & 24.78 & 1.42 & 8.82 & 22.59 & 28.17 & 2.10 & 13.34 & 27.57 & 63.60 & 25.79 & 45.42 & 53.28 \\
TSIEN &IMvMLC & & 20.85 & 0.73 & 6.17 & 19.63 & 22.50 & 0.92 & 10.35 & 20.47 & 61.92 & 12.95 & 33.89 & 52.94 \\
SIP & IMvMLC& & 24.79 & 1.02 & 8.11 & 23.21 & 24.23 & 2.18 & 13.21 & 20.15 & 61.93 & 24.92 & 38.13 & 49.92 \\
KRT &MLCIL & & 21.11 & 0.72 & 5.23 & 19.97 & 26.78 & 1.45 & 11.21 & 25.44 & 57.07 & 10.63 & 35.12 & 50.59 \\
MULTI-LANE & MLCIL& & 22.85 & 0.97 & 7.14 & 20.88 & \underline{28.41} & 2.24 & 13.61 & 27.91 & 58.98 & 11.61 & 37.09 & 52.13 \\
\midrule
AIMNet-R &IMvMLC &\multirow{4}{*}{200} & 25.59 & 1.47 & 8.82 & 23.33 & 28.40 & \underline{2.53} & \underline{13.37} & 28.34 & \underline{64.59} & 26.43 & 41.36 & 54.18 \\
SIP-R & IMvMLC& & \underline{25.89} & 1.53 & 8.81 & 24.13 & 26.61 & 1.92 & 11.87 & 25.29 & 63.98 & 26.23 & 40.11 & 52.94 \\
PRS & MLCIL & & 20.12 & 0.58 & 5.82 & 17.23 & 22.34 & 1.13 & 9.24 & 22.11 & 56.66 & 9.67 & 33.13 & 51.88 \\
KRT-R & MLCIL& & 22.56 & 0.79 & 7.21 & 21.12 & 27.42 & 1.69 & 12.19 & 26.77 & 61.01 & 16.12 & 39.98 & 52.62 \\
\midrule
AIMNet-R &IMvMLC &\multirow{4}{*}{2/class} & 25.79 & \underline{1.56} & \underline{9.84} & 24.01 & 28.03 & 2.12 & 12.34 & 27.56 & 64.29 & \underline{26.45} & \underline{46.32} & \underline{54.30} \\
SIP-R & IMvMLC& & 25.36 & 1.47 & 8.97 & \underline{24.43} & 26.72 & 1.83 & 11.13 & 25.34 & 63.71 & 25.80 & 39.50 & 52.63 \\
PRS & MLCIL & & 20.38 & 0.62 & 6.42 & 17.83 & 22.44 & 1.18 & 9.55 & 21.44& 56.65 & 9.28 & 31.98 & 50.23 \\
KRT-R & MLCIL& & 22.89 & 0.82 & 7.56 & 21.44 & 28.01 & 1.78 & 12.56 & 26.72& 60.11 & 15.34 & 39.50 & 52.77 \\
\midrule
\textsf{E2PL} & IMvMLCIL & 0 & \textbf{28.29} & \textbf{2.37} & \textbf{11.37} & \textbf{27.10} & \textbf{30.74} & \textbf{2.91}& \textbf{14.82}& \textbf{30.40} & \textbf{66.76} & \textbf{30.73} & \textbf{47.14} & \textbf{59.08} \\
\bottomrule
\end{tabular}
\label{tab:sota}
\end{table*}
% \subsection{Experimental Settings}
% This section presents a comprehensive evaluation of our proposed model, \textsf{E2PL}, against eight state-of-the-art baselines. The experimental setting is as follows.
% The experiments were conducted on the Ubuntu 22.04 system with Intel(R) Xeon(R) Gold 6326 CPU @ 2.90GHz, 256GB RAM, and NVIDIA L40 GPU (46GB memory size).
% The code is implemented in Python 3.11 with PyTorch 2.6.
\emph{Datasets.}
Following the protocol established in previous works \cite{liu2024attention}, we conduct experiments on three widely used multi-view multi-label datasets: \emph{ESPGame} \cite{von2004labeling}, \emph{IAPRTC12} \cite{grubinger2006iapr}, and \emph{MIRFLICKR} \cite{huiskes2008mir}. Each dataset is represented by six distinct feature types, including GIST, HSV, DenseHue, DenseSift, RGB, and LAB. Detailed statistics for all datasets are provided in \textbf{Appendix A}. For each dataset, we randomly choose 15\% samples for the test, 15\% for validation, and the rest for training.
\emph{Metrics.}
Following \cite{dong2023knowledge} and \cite{zhang2025l3a}, we adopt the mean average precision (mAP) as the primary evaluation metric for each incremental task, and report both the average mAP (the average of the mAP across all tasks) and the last mAP (the mAP of the
last task). We also report the per-class F1 score (CF1) and
overall F1 score (OF1) for performance evaluation.
\emph{Baselines.}
We compare \textsf{E2PL} with five state-of-the-art IMvMLC methods: DICNet~\cite{liu2023dicnet}, MTD~\cite{liu2023masked}, AIMNet~\cite{liu2024attention}, TSIEN~\cite{tan2024two}, and SIP~\cite{liu2024partial};
and three MLCIL methods: PSR~\cite{kim2020imbalanced}, KRT~\cite{dong2023knowledge}, and MULTI-LANE~\cite{de2024less}. We include the Upper-bound baseline, which assumes access to all task data simultaneously and represents the ideal performance.
Following \cite{dong2023knowledge}, we set three different memory sizes: 0, 200, and 2/class.
To ensure fair and rigorous experimental comparisons, we strictly adhere to the original implementation protocols for all MLCIL methods. As these methods are limited to single-view inputs, we independently train each method on every available view and report the best performance achieved across all views.
\emph{Implementation Details.}
Since all datasets are originally complete in terms of view availability, we artificially introduce missing data to emulate real-world incomplete scenarios.
Following~\cite{liu2023dicnet}, we generate a missing instance rate of $p\%$ (i.e., view missing rate $R_V = p\%$), by randomly removing $p\%$ of the instances from each view, while ensuring every instance remains observable in at least one view. To simulate a class-incremental learning scenario, each dataset is partitioned into $T=7$ incremental tasks, with an equal number of classes introduced at each stage. For our proposed method, the dimension $d$ of all prompts is set to 128, and all TT-ranks are set as $r_0 = r_1 = \cdots = r_n = 2$. Other hyperparameters are set as follows: $N=3$, $k=4$, $\alpha=1$, and $\lambda=0.001$. The Adam optimizer is used for all experiments, with a batch size of 128 and a learning rate of 0.02. The detailed experimental setups are provided in \textbf{Appendix B}.
For fairness, all results are averaged over five times.
% For our proposed method, the dimension $d$ of all prompts is set to 128, and all TT-ranks are set as $r_0 = r_1 = \cdots = r_n = 2$.
% Other hyperparameters are set as follows: $N=3$, $k=4$, $\alpha=1$, and $\lambda=0.001$. All experiments are conducted using the Adam optimizer with a batch size of 128 and a learning rate of 0.02. For fairness, all results are averaged over five times.
\begin{figure*}[t]
\includegraphics[width=\textwidth]{image/f3.pdf}
\caption{Comparison of performance (mAP) on three datasets at $R_V = 30\%$ over $T=7$ incremental tasks (Memory = 0).}
\label{fig:incremental}
\vspace*{-0.15in}
\end{figure*}
% \begin{table*}[ht]
% \centering
% \setlength{\tabcolsep}{3pt}
% \begin{tabular}{l c cccc cccc cccc}
% \toprule
% \multirow{3}{*}{Method} & \multirow{3}{*}{Memory}
% & \multicolumn{4}{c}{ESPGame}
% & \multicolumn{4}{c}{IAPRTC12}
% & \multicolumn{4}{c}{MIRFLICKR}\\
% \cmidrule(lr){3-6} \cmidrule(lr){7-10} \cmidrule(lr){11-14}
% & &\multicolumn{1}{c}{Avg.} & \multicolumn{3}{c}{Last}
% &\multicolumn{1}{c}{Avg.} & \multicolumn{3}{c}{Last}
% &\multicolumn{1}{c}{Avg.} & \multicolumn{3}{c}{Last}\\
% \cmidrule(lr){3-3} \cmidrule(lr){4-6} \cmidrule(lr){7-7} \cmidrule(lr){8-10} \cmidrule(lr){11-11} \cmidrule(lr){12-14}
% & & mAP & CF1 & OF1 & mAP & mAP & CF1 & OF1 & mAP & mAP & CF1 & OF1 & mAP \\
% \midrule
% Upper-bound & 0 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 \\
% \midrule
% DICNet & \multirow{5}{*}{0} & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 \\
% MTD & & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 \\
% AIMNet & & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 \\
% TSIEN & & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 \\
% SIP & & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 \\
% \midrule
% AIMNet & \multirow{4}{*}{250} & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 \\
% SIP & & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 \\
% KRT-R & & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 \\
% CSC-R & & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 \\
% \midrule
% KRT & \multirow{3}{*}{0} & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 \\
% CSC & & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 \\
% MULTI-LANE & & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 \\
% \midrule
% E2PL & 0 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 & 65.65 \\
% \bottomrule
% \end{tabular}
% \caption{Comparison with state-of-the-art methods.}
% \label{tab:sota}
% \end{table*}
\subsection{Comparison Study}
Table~\ref{tab:sota} summarizes the performance across all datasets, showing that \textsf{E2PL} consistently outperforms all baselines.
On the \emph{MIRFLICKR} dataset, under the strict zero-memory setting, \textsf{E2PL} achieves a mAP of 66.76, surpassing the best IMvMLC baseline AIMNet-R and the leading MLCIL method KRT-R by 2.47 and 5.75 points, corresponding to relative gains of 3.8\% and 9.4\%.
For the last-increment mAP, \textsf{E2PL} reaches 59.08, exceeding all baselines by 4.78 to 6.46 points. Similarly, \textsf{E2PL} achieves the highest CF1 and OF1 scores without memory replay. Consistent trends are observed on \emph{ESPGame} and \emph{IAPRTC12}, where \textsf{E2PL} achieves the best results on all metrics. Notably, \textsf{E2PL} is the only method that consistently maintains high performance under the most challenging incremental learning scenarios, underscoring its robustness and generalizability in both class-incremental and incomplete multi-view settings.
Figure~\ref{fig:incremental} provides a fine-grained depiction of mAP progression across incremental learning tasks. Unlike other methods that suffer significant drops as new classes are introduced, \textsf{E2PL} maintains stable performance, with fluctuations below 2\% across all increments. This resilience to forgetting is attributed to its prompt-based architecture, which preserves previously learned knowledge while effectively integrating new information, ensuring both stability and adaptability in class incremental learning.
\begin{figure}[t]
\centering
\includegraphics[width=1\columnwidth]{image/f4.pdf}
\caption{Performance demonstration under different view missing rates on the \emph{MIRFLICKR} dataset.}
\label{fig:mr}
\vspace*{-0.15in}
\end{figure}
\subsection{Effect of View Missing Rates}
Figure~\ref{fig:mr} illustrates the robustness of different methods under varying view missing rates ($R_V$) in the CIL setting. \textsf{E2PL} consistently achieves the highest mAP across all missing rates. At $R_V=0\%$, \textsf{E2PL} attains a mAP of 70.14, exceeding the best baseline AIMNet by 4.48 points. When $R_V$ rises to $30\%$, \textsf{E2PL} maintains a leading mAP of 66.76, while AIMNet drops to 63.60, resulting in a margin of 3.16 points. Even at $R_V=70\%$, \textsf{E2PL} achieves 60.21 mAP, outperforming the nearest competitor AIMNet by 6.13 points. Notably, as $R_V$ increases, the performance gap between \textsf{E2PL} and the baseline methods gradually narrows, decreasing from 4.48\% at $R_V=0\%$ to 2.0\% at $R_V=70\%$-which is anticipated, as higher missing rates mean less available information and greater learning difficulty, particularly in class incremental settings. Nevertheless, the prompt-based architecture of \textsf{E2PL} exhibits substantial robustness and adaptability, consistently outperforming state-of-the-art methods and ensuring reliable performance even under severe view missing scenarios.
Additional experiments on two other datasets, detailed in \textbf{Appendix C}, further confirm the effectiveness and generalizability of \textsf{E2PL}.
% \begin{table*}[h]
% \centering
% \caption{Resource consumption and computational efficiency of different methods. \textbf{Bold} and \underline{underlined} values indicate the \textbf{best} and \underline{second-best} results, respectively.}
% \label{tab:times}
% \begin{tabular}{l c c c c}
% \toprule
% \textbf{Methods} & \textbf{Params (M)} & \textbf{Training Time/Epoch (ms)} & \textbf{Inference Time/Batch (ms)} & \textbf{Max GPU Memory (MB)} \\
% \midrule
% MAP & 0.193 & \textbf{268} & \textbf{10.5} & 998 \\
% MSP & 0.049& \underline{315} & 27.6 & \underline{822} \\
% EPE-P& \underline{0.044}& 405 & 30.1 & \textbf{801} \\
% \textsf{E2PL}(Ours) &\textbf{0.026} &335 & \underline{11.4} & 889\\
% \bottomrule
% \end{tabular}
% \label{tab:Eff3}
% \end{table*}
\subsection{Analysis of Effectiveness and Efficiency}
To rigorously assess the effectiveness and efficiency of \textsf{E2PL}, we perform controlled experiments by substituting the EPT module with MAP~\cite{lee2023multimodal}, MSP~\cite{jang2024towards}, and EPE-P~\cite{chen2024epe} within a unified framework.
Here, the parameter count specifically refers to the number of prompt parameters required to model all possible missing-view patterns, where $n$ is the total number of views.
Other variables, including $d$, $r$, $R$, and $k$, are treated as constants, and all other factors are held fixed. As summarized in Table~\ref{tab:Eff}, MAP and MSP exhibit exponential parameter growth ($\mathcal{O}(2^n)$), and EPE-P shows cubic complexity, while EPT achieves linear growth ($\mathcal{O}(n)$) owing to its efficient modeling strategy.
EPT leverages atomic tensor decomposition for dimension reduction, enabling customized modeling for each missing-view pattern while simultaneously enhancing interactions and knowledge sharing across different patterns. Notably, EPT achieves the highest mAP among all compared methods, with minimal storage overhead. These results convincingly demonstrate both the effectiveness and efficiency of EPT in the IMvMLCIL setting. We further validate the superior efficiency of \textsf{E2PL} through experiments on two additional datasets in \textbf{Appendix D}.
Table~\ref{tab:Eff3} further elucidates the resource efficiency of our approach. \textsf{E2PL} requires only 0.026M parameters—an 86\% reduction compared to MAP (0.193M)—and maintains a competitive training time of 335~ms per epoch. The marginally increased training time, relative to MAP (268~ms), is primarily due to the dynamic generation of all missing-aware prompts during optimization. In the inference phase, however, \textsf{E2PL} utilizes the missing-view indicator to generate only the necessary prompt for each instance, yielding a significant advantage in inference speed (11.4~ms) over MSP (27.6~ms) and EPE-P (30.1~ms), and closely approaching MAP (10.5~ms). GPU memory consumption remains moderate at 889~MB, further highlighting the method’s practical efficiency. \textsf{E2PL} thus establishes a new paradigm for IMvMLCIL by simultaneously advancing both effectiveness and efficiency in large-scale, real-world scenarios.
\begin{table}[t]
\centering
\setlength{\tabcolsep}{5.5pt}
\caption{Analysis of effectiveness and efficiency on the \emph{MIRFLICKR} dataset at $R_V = 30\%$. \textbf{Bold} and \underline{underlined} values indicate the \textbf{best} and \underline{second-best} results, respectively.}
\begin{tabular}{lccc }
\toprule
\textbf{Methods} & \textbf{\# Params} & \textbf{Complexity} & \textbf{mAP} \\
\midrule
MAP & $(2^{n})\cdot d $ & $\mathcal{O}(2^{n})$ &65.10$\pm$0.11 \\
MSP & $n \cdot d $ & $\mathcal{O}(n)$&64.66$\pm$0.09 \\
EPE-P & $n^3 + d \cdot r $ & $\mathcal{O}(n^3)$&\underline{65.57$\pm$0.05} \\
EPT & $n \cdot R^2 \cdot k + d \cdot k $ & $\mathcal{O}(n)$& \textbf{66.76$\pm$0.03} \\
\bottomrule
\end{tabular}
\label{tab:Eff}
\end{table}
% \subsection{Parameter Evaluation}
% We perform a comprehensive sensitivity analysis on the TT-rank $R$ and factor dimension $k$ of the Efficient Prototype Tensorization module, varying both parameters within the set $\{1, 2, 4, 8\}$. As illustrated in Figure~\ref{fig:TT}, increasing $k$ and $R$ generally leads to improved mAP; however, the performance gains become marginal once $k$ and $R$ surpass moderate values. Notably, the configuration $k=4$, $R=2$ achieves the optimal balance between performance and parameter efficiency, yielding the highest mAP ($66.76\%$) with minimal parameter ($0.026 M$) overhead. This configuration strikes a crucial balance: it is expressive enough to capture diverse missing-view patterns, but remains compact to prevent overfitting and excessive resource consumption. Consequently, we adopt $k=4$, $R=2$ in our main experiments, demonstrating that \textsf{E2PL} attains robust and efficient performance with a compact parameterization, which is essential for scalable deployment in practical applications.
% \begin{figure}[t]
% \centering
% \includegraphics[width=1\columnwidth]{image/TT_Parameter_Sensitivity_Analysis.pdf}
% \caption{Parameter Sensitivity Analysis of TT-Rank $R$ and Factor Dimension $k$ on the \emph{MIRFLICKR} dataset at $R_V = 30\%$.}
% \label{fig:TT}
% \end{figure}
\subsection{Ablation Study}
To assess the effectiveness of each proposed component, we conduct an ablation study on the \emph{MIRFLICKR} dataset, as shown in Table~\ref{tab:Ablation}. Using only task-tailored prompts (TTPs) yields an mAP of 62.04, serving as the baseline for class-incremental adaptation. Introducing missing-aware prompts (MAPs) increases the mAP to 64.18, underscoring the value of explicitly modeling incomplete view configurations. Incorporating efficient prototype tensorization (EPT) further improves the mAP to 65.71; importantly, EPT not only reduces parameter complexity but also enhances interaction modeling across different missing-view scenarios, enabling more expressive and scalable prompt generation. The addition of dynamic contrastive learning (DCL) results in a further performance gain (mAP 65.38), as DCL dynamically captures the underlying relationships among various missing patterns, facilitating robust representation learning. When all modules are integrated, \textsf{E2PL} achieves the highest performance, outperforming all partial configurations. These results clearly demonstrate that each module addresses a distinct aspect of the IMvMLCIL task, and their synergistic integration is key for achieving state-of-the-art performance.
Furthermore, the effectiveness of each module in \textsf{E2PL} is substantiated by consistent ablation studies on two additional datasets, with detailed results in \textbf{Appendix E}.
\begin{table}[t]
\centering
\caption{Resource consumption and computational efficiency of different methods. \textbf{Bold} and \underline{underlined} values indicate the \textbf{best} and \underline{second-best} results, respectively. TT/E refers to the training time per epoch for the model, IT/B denotes the inference time for a single batch sample, and MGM indicates the maximum GPU memory usage recorded during the experimental process.}
\setlength{\tabcolsep}{3pt}
\label{tab:times}
\begin{tabular}{l c c c c}
\toprule
\textbf{Methods} & \textbf{Params (M)} & \textbf{TT/E (ms)} & \textbf{IT/B (ms)} & \textbf{MGM (MB)} \\
\midrule
MAP & 0.193 & \textbf{268} & \textbf{10.5} & 998 \\
MSP & 0.049& \underline{315} & 27.6 & \underline{822} \\
EPE-P& \underline{0.044}& 405 & 30.1 & \textbf{801} \\
\textsf{E2PL} &\textbf{0.026} &335 & \underline{11.4} & 889\\
\bottomrule
\end{tabular}
\label{tab:Eff3}
\end{table}
\begin{table}[t]
\centering
\setlength{\tabcolsep}{8.3pt}
\newcommand{\cmark}{\checkmark}
\caption{Ablation results of \textsf{E2PT} on the \emph{MIRFLICKR} dataset at $R_V = 30\%$. \textbf{Bold} and \underline{underlined} values indicate the \textbf{best} and \underline{second-best} results, respectively.}
\begin{tabular}{ccccc}
\toprule
\multicolumn{4}{c}{\textbf{Components}} & \multirow{2}{*}{\textbf{mAP}}\\
\cmidrule(lr){1-4}
\textbf{+TTPs} & \textbf{+MAPs} & \textbf{+EPT} & \textbf{+DCL} & \\
\midrule
& \cmark & \cmark & \cmark & 62.04$\pm$0.11 \\
\cmark & & & & 64.18$\pm$0.05\\
\cmark & \cmark & & \cmark & \underline{65.71$\pm$0.08}\\
\cmark & \cmark & \cmark & & 65.38$\pm$0.07\\
\cmark & \cmark & \cmark & \cmark & \textbf{66.76$\pm$0.03}\\
\bottomrule
\end{tabular}
\label{tab:Ablation}
\vspace*{-0.15in}
\end{table}
\section{Conclusion}
\label{sec:Conclusions}
In this work, we propose \textsf{E2PL}, an effective and efficient prompt learning framework that integrates TTPs for class-incremental adaptation with MSPs for flexible handling of missing views.
Our EPT module reduces prompt space complexity from exponential to linear and enables expressive modeling across diverse missing-view scenarios.
Additionally, the DCL strategy further enhances robustness by capturing latent relationships among missing patterns.
Extensive experiments on three benchmarks demonstrate that \textsf{E2PL} consistently outperforms existing methods, highlighting its effectiveness and scalability in incremental scenarios.
% This paper investigates the IMvMLCIL task, requiring joint adaptation to incomplete multi-view data and evolving class spaces. We introduce \textsf{E2PL}, which combines Task-Tailored Prompts and Missing-Aware Prompts within a unified framework. The Efficient Prototype Tensorization module fundamentally alleviates the parameter explosion issue associated with missing-view modeling, while the Dynamic Contrastive Learning component strengthens the model’s robustness across diverse missing-view patterns. Empirical results on multiple benchmarks verify that \textsf{E2PL} consistently achieves state-of-the-art effectiveness and efficiency, establishing its scalability and reliability for complex, real-world incremental learning applications.
\section*{Acknowledgments}
This work is supported by the Leading Goose R\&D Program of Zhejiang under Grant No. 2024C01109, the Zhejiang Provincial Natural Science Foundation of China under Grant No.QN26F020098, the Young Doctoral Innovation Research Program of Ningbo Natural Science Foundation Grant No. 2024J207, and the Ningbo Yongjiang Talent Program Grant 2024A-158-G.
Yangyang Wu is the corresponding author of the work.
\section{Dataset Statistics and Descriptions}
For our experiments, we selected three popular benchmark datasets: \emph{ESPGame} \cite{von2004labeling}, \emph{IAPRTC12} \cite{grubinger2006iapr}, and \emph{MIRFLICKR} \cite{huiskes2008mir}.
All datasets consist of six distinct views, including GIST, HSV, Hue,
Sift, RGB, and LAB. Table \ref{tab:datasets} presents an overview of the all datasets, including the number of samples (\# Samples), the number of labels (\# Labels), and average labels per sample (\# Label Density). The details of each dataset are described as follows:
\begin{itemize}
\item \emph{ESPGame}: The \emph{ESPGame} dataset consists of 20770 images and 20 categories, and each image can be associated
with multiple tags.
\item \emph{IAPRTC12}: The \emph{IAPRTC12} dataset contains 20000 images from the IAPR project, annotated with multiple categories such as indoor and outdoor scenes, objects, and animals. In our experiments, we choose 19627 samples with 291 categories annotated.
\item \emph{MIRFLICKR}: The \emph{MIRFLICKR} dataset includes 25000 images from the Flickr platform and includes 38 tags.
\end{itemize}
\begin{table}[t]
\centering
\caption{Statistics of all datasets used in our experiments.}
\begin{tabular}{lccccc}
\toprule
Datasets & \# Samples & \# Labels & \# Label Density \\
\midrule
\textit{ESPGame} & 20770 & 268 & 4.686 \\
\textit{IAPRTC12} & 19627 & 291 & 5.719 \\
\textit{MIRFLICKR} & 25000 & 38 & 4.716 \\
\bottomrule
\end{tabular}
\label{tab:datasets}
\end{table}
\begin{table}[h]
\centering
\setlength{\tabcolsep}{1pt}
\caption{Analysis of efficiency on the \emph{ESPGame} and \emph{IAPRTC12} datasets at $R_V = 30\%$. \textbf{Bold} and \underline{underlined} values indicate the \textbf{best} and \underline{second-best} results, respectively.}
\begin{tabular}{l c c c c}
\toprule
\multirow{2}{*}{\textbf{Methods}} & \multirow{2}{*}{\textbf{\# Params}} & \multirow{2}{*}{\textbf{Complexity}} & \multicolumn{2}{c}{\textbf{mAP}} \\
\cmidrule(lr){4-5}
& & & \emph{ESPGame} & \emph{IAPRTC12} \\
\midrule
MAP & $(2^n) \cdot d$ & $\mathcal{O}(2^n)$ & 27.12 $\pm$ 0.08 & 29.78 $\pm$ 0.02 \\
MSP & $n \cdot d$ & $\mathcal{O}(n)$ & 26.78 $\pm$ 0.01 & 29.19 $\pm$ 0.04 \\
EPE-P & $n^3 + d \cdot r$ & $\mathcal{O}(n^3)$ & \underline{27.44 $\pm$ 0.03} & \underline{29.90 $\pm$ 0.04} \\
EPT & $n \cdot R^2 \cdot k + d \cdot k$ & $\mathcal{O}(n)$ & \textbf{28.29 $\pm$ 0.02} & \textbf{30.74 $\pm$ 0.04} \\
\bottomrule
\end{tabular}
\label{tab:Eff2}
\end{table}
\section{Experimental Setup}
% This section presents a comprehensive evaluation of our proposed model, \textsf{E2PL}, against eight state-of-the-art baselines. The experimental setting is as follows.
The experiments were conducted on the Ubuntu 22.04 system with Intel(R) Xeon(R) Gold 6326 CPU @ 2.90GHz, 256GB RAM, and NVIDIA L40 GPU (46GB memory size).
\emph{Class Incremental Task Setup.}
We evaluate our method under the \emph{incomplete multi-view multi-label class incremental learning} (IMvMLCIL) setting, as defined above. The dataset is organized into $T$ incremental sessions $\{ (\mathcal{X}^1, \mathcal{Y}^1, \mathcal{C}^1), \cdots, (\mathcal{X}^T, \mathcal{Y}^T, \mathcal{C}^T) \}$, where each session $t$ introduces a disjoint class set $\mathcal{C}^t$ ($\mathcal{C}^m \cap \mathcal{C}^n = \emptyset$ for $m \neq n$). Given $C$ total classes, the first session contains $C_{\text{base}}$ base classes, and the remaining $C - C_{\text{base}}$ classes are evenly divided among the subsequent $T-1$ sessions. The number of new classes per incremental session is
\begin{equation}
C_{\text{inc}} = \frac{C - C_{\text{base}}}{T - 1},
\end{equation}
where $C_{\text{inc}}$ is required to be an integer for balanced partitioning.
In the multi-label scenario, each sample may be annotated with multiple labels spanning different sessions. To ensure all relevant supervisory signals are available, a sample $x_i$ is included in session $t$ if its label set overlaps with $\mathcal{C}^t$, i.e.,
\begin{equation}
x_i \in \mathcal{X}^t \iff \text{Label}(x_i) \cap \mathcal{C}^t \neq \emptyset.
\end{equation}
This protocol allows the model to observe samples containing both current and previously seen classes, which is essential for mitigating catastrophic forgetting in MvMLCIL. Notably, this may result in the same sample appearing in multiple sessions—a natural property of multi-label incremental learning.
For all datasets, we set $T = 7$ and adopt the following configurations: \emph{ESPGame} ($C_{\text{base}} = 208$, $C_{\text{inc}} = 10$), \emph{IAPRTC12} ($C_{\text{base}} = 231$, $C_{\text{inc}} = 10$), and \emph{MIRFLICKR} ($C_{\text{base}} = 7$, $C_{\text{inc}} = 8$). This setup enables a rigorous evaluation of model plasticity and stability under realistic, complex IMvMLCIL conditions.
\section{Effect of View Missing Rates}
To further validate the robustness of \textsf{E2PL} under varying degrees of view incompleteness, we conduct additional experiments on the \emph{ESPGame} and \emph{IAPRTC12} datasets, systematically increasing the view missing rate $R_V$ from 0\% to 70\%. Figure~\ref{fig:mr2} and Figure~\ref{fig:mr3} present the mean Average Precision (mAP) of \textsf{E2PL} and several state-of-the-art baselines under different $R_V$ settings.
On the \emph{ESPGame} dataset, \textsf{\textsf{E2PL}} consistently achieves the highest mAP across all missing rates. For example, at $R_V=0\%$, \textsf{\textsf{E2PL}} obtains an mAP of 31.56, outperforming the best baseline (MULTI-LANE, 27.67) by 3.89 points. As the missing rate increases to 30\%, \textsf{\textsf{E2PL}} maintains a leading mAP of 28.29, which is 5.44 points higher than both AIMNet and MULTI-LANE (22.85). Even under the most challenging scenario with $R_V=70\%$, \textsf{\textsf{E2PL}} achieves 22.75 mAP, still surpassing AIMNet (21.27) by 1.48 points. A similar trend is observed on the \emph{IAPRTC12} dataset. \textsf{\textsf{E2PL}} achieves an mAP of 35.98 at $R_V=0\%$, which is 3.95 points higher than MULTI-LANE (32.03). As $R_V$ increases to 30\%, \textsf{\textsf{E2PL}} maintains a strong lead with 30.74 mAP, compared to 28.41 for MULTI-LANE and 28.17 for AIMNet. Even at $R_V=70\%$, \textsf{\textsf{E2PL}} achieves 25.44 mAP, outperforming all baselines by at least 2.53 points.
\begin{figure}[h]
\centering
\includegraphics[width=\columnwidth]{image/f5.pdf}
\caption{Performance demonstration under different view missing rates on the \emph{IAPRTC12} dataset.}
\label{fig:mr3}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=\columnwidth]{image/f6.pdf}
\caption{Performance demonstration under different view missing rates on the \emph{ESPGame} dataset.}
\label{fig:mr2}
\end{figure}
These findings are highly consistent with the results reported in the main text (see Section 5.2), where \textsf{\textsf{E2PL}} demonstrates remarkable robustness and stability under increasing view missing rates on the \emph{MIRFLICKR} dataset. Across all three benchmarks, \textsf{\textsf{E2PL}} exhibits a significantly slower performance degradation compared to existing methods, confirming the effectiveness of our efficient prototype tensorization and dynamic contrastive learning strategies. This strong and consistent advantage highlights the generalizability and practical value of \textsf{E2PL} for real-world incomplete multi-view learning scenarios.
\begin{table}[h]
\centering
\setlength{\tabcolsep}{3.5pt}
\caption{Ablation results of \textsf{E2PT} on the \emph{ESPGame} and \emph{IAPRTC12} datasets at $R_V = 30\%$. \textbf{Bold} and \underline{underlined} values indicate the \textbf{best} and \underline{second-best} results, respectively.}
\newcommand{\cmark}{\checkmark}
\begin{tabular}{cccccc}
\toprule
\multicolumn{4}{c}{\textbf{Component}} & \multicolumn{2}{c}{\textbf{mAP}} \\
\cmidrule(lr){1-4} \cmidrule(lr){5-6}
\textbf{+TTPs} & \textbf{+MAPs} & \textbf{+EPT} & \textbf{+DCL} & \emph{ESPGame} & \emph{IAPRTC12} \\
\midrule
& \cmark & \cmark & \cmark & 23.87 $\pm$ 0.11 & 24.12 $\pm$ 0.06 \\
\cmark & & & & 26.23 $\pm$ 0.06 & 28.12 $\pm$ 0.03\\
\cmark & \cmark & & \cmark & \underline{27.55 $\pm$ 0.03} & \underline{29.62 $\pm$ 0.05} \\
\cmark & \cmark & \cmark & & 27.40 $\pm$ 0.02 & 29.37 $\pm$ 0.04 \\
\cmark & \cmark & \cmark & \cmark & \textbf{28.29 $\pm$ 0.02} & \textbf{30.74 $\pm$ 0.04} \\
\bottomrule
\end{tabular}
\label{tab:Ablation2}
\end{table}
\section{Efficiency Analysis}
To rigorously evaluate the effectiveness and efficiency of \textsf{E2PL}, we conduct controlled experiments by replacing the Efficient Prototype Tensorization (EPT) module with MAP~\cite{lee2023multimodal}, MSP~\cite{jang2024towards}, and EPE-P~\cite{chen2024epe} within a unified framework, thereby enabling a direct and systematic comparison of prompt parameterization strategies in terms of predictive performance and parameter complexity.
As reported in Table~\ref{tab:Eff2}, \textsf{E2PL} consistently achieves the highest mAP on both \emph{ESPGame} (28.29) and \emph{IAPRTC12} (30.74), substantially outperforming MAP (27.12/29.78), MSP (26.78/29.19), and EPE-P (27.44/29.90). Importantly, \textsf{E2PL} attains these results with only $\mathcal{O}(n)$ parameter complexity, in stark contrast to the exponential growth of MAP and the cubic complexity of EPE-P. This linear scalability, enabled by the EPT module, allows for expressive and efficient modeling of arbitrary missing-view patterns while preserving computational tractability.
\section{Ablation Studies}
To further validate the necessity and effectiveness of each module in \textsf{E2PL}, we conduct comprehensive ablation studies on the \emph{ESPGame} and \emph{IAPRTC12} datasets, as shown in Table~\ref{tab:Ablation2}. Consistent with the main text, we observe that each component contributes distinct and complementary improvements. Starting from the baseline with only task-tailored prompts (TTPs), the introduction of missing-aware prompts (MAPs) consistently yields a notable boost in mAP, with relative improvements exceeding 10\% on both datasets. Incorporating the efficient prototype tensorization (EPT) module not only further enhances performance, but also dramatically reduces parameter complexity, enabling scalable prompt generation for arbitrary missing-view patterns. The addition of dynamic contrastive learning (DCL) brings further, albeit more moderate, gains, particularly in challenging scenarios with severe view incompleteness. When all modules are integrated, \textsf{E2PL} achieves the best overall results, outperforming all ablated variants by a clear margin. These trends are fully aligned with those observed on \emph{MIRFLICKR} in the main paper, underscoring the robustness and generalizability of our approach. Collectively, these findings confirm that the synergy among TTPs, MAPs, EPT, and DCL is essential for achieving state-of-the-art performance in the IMvMLCIL setting.
\begin{figure}[t]
\centering
\includegraphics[width=1\columnwidth]{image/TT_Parameter_Sensitivity_Analysis.pdf}
\caption{Parameter Sensitivity Analysis of TT-Rank $R$ and Factor Dimension $k$ on the \emph{MIRFLICKR} dataset at $R_V = 30\%$.}
\label{fig:TT}
\end{figure}
\section{Parameter Evaluation}
We perform a comprehensive sensitivity analysis on the TT-rank $R$ and factor dimension $k$ of the Efficient Prototype Tensorization module, varying both parameters within the set $\{1, 2, 4, 8\}$. As illustrated in Figure~\ref{fig:TT}, increasing $k$ and $R$ generally leads to improved mAP; however, the performance gains become marginal once $k$ and $R$ surpass moderate values. Notably, the configuration $k=4$, $R=2$ achieves the optimal balance between performance and parameter efficiency, yielding the highest mAP ($66.76\%$) with minimal parameter ($0.026 M$) overhead. This configuration strikes a crucial balance: it is expressive enough to capture diverse missing-view patterns, but remains compact to prevent overfitting and excessive resource consumption. Consequently, we adopt $k=4$, $R=2$ in our main experiments, demonstrating that \textsf{E2PL} attains robust and efficient performance with a compact parameterization, which is essential for scalable deployment in practical applications.