Enhancing explainability via distinguishable feature learning based on causality in image classification

Ning Yu, Long Chen, Xiaoyin Yi

Abstract: Highlights•The features are decomposed through the gradient orthogonality into causal features with a strong correlation with the predicted category and non-causal features related to the contextual information. And, we use causality-based features as sufficient learning conditions.•Focuses on clustering causal features within the same category, rather than clustering all features of the same category.•With the SHAP (SHapley Additive exPlanations), we calculated feature attribution for the model behavior to generate a feature attribution map, highlighted regions with higher values contributing to the categories.