Adaptive Subgraph Feature Extraction for Explainable Multi-Modal Learning

This work introduces a pioneering methodology aimed at advancing explainability within Graph Neural Networks (GNNs) across multi-modal data domains. Recognizing the intricate nature of GNNs and the challenge posed by their inherent "black-box" characteristics, especially in multimodal learning environments, our approach seeks to demystify these complex decision-making processes. Leveraging Adaptive Subgraph Feature Extraction, we propose a novel framework that dynamically identifies and extracts significant subgraph features contributing to the GNN’s decision-making. This methodology not only illuminates the path to understanding GNN operations but also ensures model transparency without compromising performance. Empirical validation on the comprehensive Amazon dataset demonstrates our approach’s effectiveness, significantly outperforming existing baselines in fidelity, sparsity, validity, and consistency metrics. Our contributions promise a substantial leap towards transparent, explainable, and reliable GNNs in multimodal learning contexts, paving the way for future explorations aimed at refining and expanding the horizons of explainable AI.