ICFI: A Feature Importance Measure For Multi-Class Classification

Tommaso Amico, Pernille Matthews, Ira Assent

Keywords: XAI, Explainaility, Feature Importance, Multi-class Classification, Classification

TL;DR: We propose a feature importance measure highlighting the features the model uses to separate two classes

Abstract: Feature importance is one of the most prominent methods in eXplainable Artificial Intelligence (XAI). It aims to assess the extent to which a machine learning model relies on different features. However, in multi-class classification, current methods fail to explain inter-class relationships, either because they provide explanations for binary classification only, or because they suffer from aggregation bias. To address these shortcomings, we propose Inter-Class Feature Importance (ICFI), which provides feature importance scores for discriminating between an arbitrary pair of classes. ICFI is a post-hoc, model-agnostic method, which provides bounded scores for interpretability. We empirically demonstrate through extensive experiments on real-world datasets that ICFI effectively captures the discriminating features between class pairs, outperforming existing methods.

Primary Area: interpretability and explainable AI

Submission Number: 11089