New AI Handles Incomplete Multimodal Sentiment Analysis

A new research paper introduces MIDAS (Mutual Information Disentanglement with uncertainty-Aware fuSion), a novel framework designed to address the significant challenge of incomplete or corrupted multimodal inputs in sentiment analysis. Traditional multimodal approaches often assume complete data, which is rarely the case in real-world applications. MIDAS tackles this by restructuring multimodal representations under these challenging conditions. The framework employs a variational modeling strategy, representing each modality with multivariate Gaussian latent variables. These variables are then decomposed into shared and exclusive factors. To ensure reliable representations, MIDAS uses a minimax objective that minimizes mutual information between shared and exclusive spaces for stable disentanglement, while maximizing mutual information among shared spaces across modalities to enhance semantic alignment. Furthermore, MIDAS incorporates an uncertainty-aware fusion mechanism. This mechanism leverages posterior variance as a reliability indicator, adaptively weighting latent features during the fusion process. This ensures robust integration even when modalities are incomplete, leading to strong and consistent performance gains over competitive baselines across various incomplete data settings on widely used datasets.