Structure-Aware Prototype Guided Trusted Multi-View Classification

TL;DR

This work tackles trustworthy multi-view classification under conflicting and heterogeneous views by introducing a structure-aware prototype framework. It builds a two-stage approach: learning structure-aware prototypes per view and performing prototype-guided fusion with Evidential Deep Learning to model uncertainty. The Prototype-Guided Fine-Grained Fusion uses class-level prototypes, view Belief, and Prototype Correlation to adaptively weight views and fuse evidence, avoiding costly explicit graphs. Across six public multi-view datasets, the method achieves competitive accuracy and superior robustness to view conflicts, while maintaining computational efficiency comparable to graph-free baselines. The framework offers practical, uncertainty-aware TMVC with scalable learning dynamics for real-world, heterogeneous data fusion tasks.

Abstract

Trustworthy multi-view classification (TMVC) addresses the challenge of achieving reliable decision-making in complex scenarios where multi-source information is heterogeneous, inconsistent, or even conflicting. Existing TMVC approaches predominantly rely on globally dense neighbor relationships to model intra-view dependencies, leading to high computational costs and an inability to directly ensure consistency across inter-view relationships. Furthermore, these methods typically aggregate evidence from different views through manually assigned weights, lacking guarantees that the learned multi-view neighbor structures are consistent within the class space, thus undermining the trustworthiness of classification outcomes. To overcome these limitations, we propose a novel TMVC framework that introduces prototypes to represent the neighbor structures of each view. By simplifying the learning of intra-view neighbor relations and enabling dynamic alignment of intra- and inter-view structure, our approach facilitates more efficient and consistent discovery of cross-view consensus. Extensive experiments on multiple public multi-view datasets demonstrate that our method achieves competitive downstream performance and robustness compared to prevalent TMVC methods.

Figures (8)

• Figure 1: Comparison illustration. (a) Traditional TMVC method uses DNNs $\{f^{(m)}(\cdot)\}^{n}_{m=1}$ for evidence extraction but ignores latent neighborhood structures, risking unreliability with conflicting views. (b) TUEND adds feature-neighborhood info via DNNs $\{f^{(m)}(\cdot)\}^{n}_{m=1}$ and GNNs $\{g^{(m)}(\cdot)\}^{n}_{m=1}$, but explicit graph construction raises complexity and limits scalability. (c) Our method focuses on local prototype structures, integrating latent neighborhood structures flexibly in a computation-friendly manner.
• Figure 2: Overview of the proposed framework. The process begins by transforming multi-view data into structured feature representations using view-specific deep neural networks. Evidence is then derived from both view-specific outputs and prototype-based embeddings that dynamically encode intra- and inter-view relations. These sources of evidence are subsequently integrated through Prototype-Guided Fine-Grained Fusion (PFF), yielding robust predictions along with quantified uncertainty.
• Figure 3: Effect of prototype correlation values
• Figure 4: Illustration of the relationship between prototype-derived evidence uncertainty and prediction correctness on three datasets.
• Figure 5: Visualization of evidence distributions across six views and the fused joint evidence.