Beyond Positive and Negative: A Directional SHAP Framework for Sequential Multi-Class Classification

Abstract

In this study, we address the interpretative limitations of the standard Shapley Additive Explanations (SHAP) method in sequential multi-class classification problems within the scope of Explainable Artificial Intelligence (XAI). Stemming from the observation that classical SHAP is restricted to revealing only positive and negative contributions for a single class, we propose a novel directional framework that categorizes feature effects as 'Lower' (driving towards a lower class), 'Upper' (driving towards a higher class), and 'Ambiguous' (representing inconsistent effects). To validate this approach, a Random Forest model predicting obesity levels across seven hierarchical classes was trained on an open-source dataset, achieving a classification accuracy of 95.5%. Furthermore, a stability analysis comprising 10,000 sampling iterations demonstrated the robustness of the proposed framework, with dominant features retaining their directional categorizations consistently in over 99.9% of the trials. The findings indicate that unlike standard SHAP, our method successfully isolates the specific variables that prevent an instance from ascending to a higher class or descending to a lower one, particularly clarifying the role of ambiguous boundary features. In conclusion, this modification significantly enhances model transparency for complex hierarchical scenarios, and the framework has been released as an open-source Python library to provide researchers with a practical tool for automated directional feature analysis.