Feature Attribution Explanation Based on Multi-Objective Evolutionary Learning

Feature attribution explanation (FAE) method, which reveals the contribution of each input feature to the model's output, is one of the most popular explainable artificial intelligence techniques. To assess the quality of explanations provided by FAE methods, various metrics spanning multiple dimensions have been introduced. However, current FAE approaches often prioritize faithfulness of their explanations, neglecting other crucial aspects. To address this issue, we define the construction of a FAE explainable model as a multi-objective learning problem and propose a framework that simultaneously considers multiple quality metrics during FAE explanation generation. Our experimental results demonstrate that our approach outperforms existing FAE methods in terms of faithfulness, sensitivity, and complexity. Moreover, our method has better diversity and the capacity to offer different explanations for different stakeholders.This paper for the Hot-off-the-Press track at GECCO 2024 summarizes the work Ziming Wang, Changwu Huang, Yun Li, and Xin Yao: Multi-objective Feature Attribution Explanation for Explainable Machine Learning published in ACM Transactions on Evolutionary Learning and Optimization [5].