Evidential Deep Partial Multi-View Classification With Discount Fusion

TL;DR

This work tackles incomplete multi-view classification by combining uncertainty-aware evidence with conflict-aware fusion. It introduces EDP-MVC, which uses K-means imputation to complete missing views and a CAEFN that learns discount factors to modulate cross-view evidence fusion, mitigating conflicts and unreliable imputations. The method relies on evidential deep learning to produce Dirichlet-distributed evidence per view and employs a discount-based aggregation, with a loss that jointly optimizes accuracy and calibrated uncertainty. Empirical results across diverse benchmarks demonstrate competitive or superior performance, especially under high missing rates and conflicting data, highlighting the practical value of reliable evidential fusion in incomplete multi-view problems.

Abstract

Incomplete multi-view data classification poses significant challenges due to the common issue of missing views in real-world scenarios. Despite advancements, existing methods often fail to provide reliable predictions, largely due to the uncertainty of missing views and the inconsistent quality of imputed data. To tackle these problems, we propose a novel framework called Evidential Deep Partial Multi-View Classification (EDP-MVC). Initially, we use K-means imputation to address missing views, creating a complete set of multi-view data. However, the potential conflicts and uncertainties within this imputed data can affect the reliability of downstream inferences. To manage this, we introduce a Conflict-Aware Evidential Fusion Network (CAEFN), which dynamically adjusts based on the reliability of the evidence, ensuring trustworthy discount fusion and producing reliable inference outcomes. Comprehensive experiments on various benchmark datasets reveal EDP-MVC not only matches but often surpasses the performance of state-of-the-art methods.

Figures (3)

• Figure 1: Illustration of the proposed EDP-MVC. Initially, data with some missing views are identified and imputed to form complete multi-view data. Each view is then processed to extract evidence, including belief masses and uncertainties. The extracted evidence is evaluated for reliability, identifying conflicting, consistent, and unreliable views. To avoid unreliable fusion results and subsequent degradation, learnable weights, representing the reliability of each view, are used as discount factors for dynamic and adaptive evidence fusion. Through this dynamic reliability-driven discount fusion, the accuracy of uncertainty quantification is enhanced, and the reliability of decision-making is improved.
• Figure 2: Classification performance with varied missing rates. The results demonstrate that our proposed method (Ours) consistently outperforms other cutting-edge methods (UIMC, DeepIMV, DCP, and CPM-Nets) on large-scale datasets, including Caltech-101 and NUS-WIDE-OBJECT. This superior performance highlights the robustness and effectiveness of our approach in handling large and complex incomplete datasets.
• Figure 3: Parameter evaluation for $\lambda_{t} = \min(\mathcal{Q}, t/\mathcal{P})$ on BRCA. The figures show the impact of the annealing step ($\mathcal{P}$ in panel (a)) and the converged value ($\mathcal{Q}$ in panel (b)) on inference accuracy under different missing rates ($\eta = 0, 0.2, 0.4$). Both parameters affect the inference accuracy within a certain range, demonstrating their significance in the proposed method.

Theorems & Definitions (3)

• Definition 1
• Definition 2
• Definition 3