UVaT: Uncertainty Incorporated View-Aware Transformer for Robust Multi-View Classification.

Area of Science:

• Computer Science
• Machine Learning
• Data Science

Background:

• Multi-view classification algorithms often assume complete and clean data across all views.
• Real-world data frequently exhibit missing representations or noise in some views, degrading algorithm performance.
• Existing methods typically address missing or noisy views separately, failing when both issues coexist.

Purpose of the Study:

• To develop a unified multi-view classification framework robust to both incomplete and noisy views.
• To create a flexible model capable of adaptively identifying view and feature significance.
• To address varying view missing patterns between training and testing phases.

Main Methods:

• Integrated early and late fusion techniques within a single framework.
• Employed a view-aware transformer in the early fusion module to handle missing views and explore inter-view relationships.
• Incorporated single-view classification and category-consistency constraints to mitigate reliance on specific view-missing patterns.
• Quantified view uncertainty in an ensemble manner for the late fusion module to estimate noise levels and enhance robustness.

Main Results:

• The proposed framework demonstrates simultaneous robustness to both incomplete and noisy views.
• Experimental results on diverse datasets confirm the model's effectiveness.
• The end-to-end trained framework successfully integrates early and late fusion strategies.

Conclusions:

• The novel framework offers a significant advancement in multi-view classification by handling data imperfections.
• The approach provides a flexible and adaptive solution for real-world datasets with missing or noisy views.
• The model's robustness and effectiveness are validated through comprehensive experiments.