A Structured Guide to Selecting Feature Attribution Techniques

Understanding the decision-making process of machine learning models has driven significant advances in explainable artificial intelligence (XAI). Feature attribution methods play a central role in explaining model behavior, yet selecting an appropriate method remains challenging due to the wide range of available techniques and the lack of standardized evaluation metrics. Without clear selection criteria, practitioners risk choosing suboptimal or misleading explanations. This survey addresses this challenge by providing a structured synthesis of the evaluation landscape. We introduce a multi-layer framework that organizes evaluation metrics into model-centric, explanation-centric, and human-centric layers, providing a systematic perspective for assessing feature attribution methods. We present practical guidelines to support practitioners in selecting appropriate methods based on this framework, while acknowledging the inherent context-dependency of method performance. We also identify open challenges in the field, including metric inconsistencies, trade-offs between explanation properties, and the integration of computational and human-centered evaluations. Our findings underscore the need for more formalized selection strategies and practical deployment frameworks, bridging the gap between theoretical advancements and real-world applications.

In The World Conference on Explainable Artificial Intelligence