E2PL: Effective and Efficient Prompt Learning for Incomplete Multi-view Multi-Label Class Incremental Learning

TL;DR

The paper tackles incomplete multi-view multi-label class incremental learning (IMvMLCIL), where data are sporadically available across views and new classes appear over time. It introduces E2PL, a prompt-learning framework that combines task-tailored prompts for class incremental adaptation with missing-aware prompts for arbitrary view-missing scenarios, supported by Efficient Prototype Tensorization (EPT) and Dynamic Contrastive Learning (DCL). EPT reduces the otherwise exponential prompt-parameter growth to linear in the number of views via atomic tensor decomposition and TT blocks, while DCL captures relationships among missing-view patterns to improve robustness. Empirical results on ESPGame, IAPRTC12, and MIRFLICKR show that E2PL achieves state-of-the-art performance with significantly reduced parameter counts and competitive inference times, demonstrating strong effectiveness and scalability for real-world web-scale settings.

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.

Figures (7)

• Figure 1: (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.
• Figure 2: 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.
• Figure 3: Comparison of performance (mAP) on three datasets at $R_V = 30\%$ over $T=7$ incremental tasks (Memory = 0).
• Figure 4: Performance demonstration under different view missing rates on the MIRFLICKR dataset.
• Figure 5: Performance demonstration under different view missing rates on the IAPRTC12 dataset.