To the Best of Trust: Full-Stage Trusted Multi-modal Clustering

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Multi-modal clustering integrates information from diverse sources to reveal underlying structures and improve performance.
• Existing methods often overlook data (aleatoric) and model (epistemic) uncertainties, focusing primarily on predictive uncertainty.
• This limitation can hinder robustness and the reliability of learned representations and clustering outcomes.

Purpose of the Study:

• To propose a novel Full-Stage Trusted Multi-modal Clustering (FSTMC) method.
• To jointly leverage aleatoric, epistemic, and predictive uncertainties for enhanced clustering.
• To develop more reliable feature representations and achieve superior clustering accuracy.

Main Methods:

• Representation learning utilizes probabilistic modeling for stable latent representations and aleatoric uncertainty estimation.
• Epistemic uncertainty is estimated using structured random perturbations during representation learning.
• An evidence-based fusion strategy with Dempster-Shafer theory is employed for dynamic multi-modal clustering, mitigating overconfidence and conflicts via uncertainty-guided prior constraints.

Main Results:

• The FSTMC method effectively integrates aleatoric, epistemic, and predictive uncertainties.
• Probabilistic modeling and evidence-based fusion lead to calibrated predictive uncertainty and reliable feature representations.
• Benchmark experiments show significant accuracy improvements over state-of-the-art multi-modal clustering techniques.

Conclusions:

• The proposed FSTMC method offers a robust framework for multi-modal clustering by comprehensively addressing various uncertainty types.
• Jointly considering aleatoric, epistemic, and predictive uncertainties leads to more accurate and reliable clustering results.
• This approach advances the field by providing a more trustworthy and effective method for integrating multi-modal data.