Explainable Multimodal Regression with Information Decomposition

Published:Dec 26, 2025 18:07
1 min read
ArXiv

Analysis

This paper addresses the interpretability problem in multimodal regression, a common challenge in machine learning. By leveraging Partial Information Decomposition (PID) and introducing Gaussianity constraints, the authors provide a novel framework to quantify the contributions of each modality and their interactions. This is significant because it allows for a better understanding of how different data sources contribute to the final prediction, leading to more trustworthy and potentially more efficient models. The use of PID and the analytical solutions for its components are key contributions. The paper's focus on interpretability and the availability of code are also positive aspects.

Reference

The framework outperforms state-of-the-art methods in both predictive accuracy and interpretability.