Beyond Heatmaps: Unsupervised Concept-Graph Reasoning for Interpretable Visual Explanation

Interpretability without predefined concepts: G-CBM automatically discovers concepts from images using unsupervised learning (Non-negative Matrix Factorization) rather than requiring humans to manually define which concepts the model should use, making it more flexible and potentially less biased by human assumptions about what matters.

Spatial grounding improves transparency: Unlike traditional explanation methods that provide single scores per concept, G-CBM identifies where concepts appear in images and how often they recur, enabling users to verify whether the model is actually looking at relevant image regions when making decisions.

Concept relationships matter: The model uses a Graph Attention Network to capture how concepts depend on and influence each other, moving beyond treating concepts in isolation—this better reflects how humans reason about complex visual decisions.

Selective concept use improves performance: The research found that filtering out weak concept signals actually improved accuracy while reducing the number of concepts needed (e.g., achieving 96% accuracy on skin lesion classification using only 2 of 10 concepts), suggesting that interpretability and performance can align rather than trade off against each other.

Competitive with supervised methods without manual annotation: G-CBM matched the performance of approaches requiring expensive human-labeled concept annotations on medical imaging tasks, suggesting unsupervised concept discovery could reduce the annotation burden in developing interpretable AI systems for specialized domains.