Concept-based Explainable Artificial Intelligence: A Survey

TL;DR

This survey addresses the need for human-understandable AI explanations by focusing on Concept-based XAI (C-XAI). It defines concepts and concept-based explanations, and introduces a nine-category taxonomy that organizes post-hoc and explainable-by-design approaches across symbolic, unsupervised, prototype, and textual concept types. The paper provides a comprehensive evaluation framework, including metrics, human studies, and datasets, and highlights initial applications and promising research directions. Its insights aim to guide researchers, practitioners, and policymakers toward more reliable, interpretable AI through concept-centric explanations that align with human reasoning. Overall, the work offers a foundational reference to advance, compare, and deploy concept-based explanations in diverse domains.

Abstract

The field of explainable artificial intelligence emerged in response to the growing need for more transparent and reliable models. However, using raw features to provide explanations has been disputed in several works lately, advocating for more user-understandable explanations. To address this issue, a wide range of papers proposing Concept-based eXplainable Artificial Intelligence (C-XAI) methods have arisen in recent years. Nevertheless, a unified categorization and precise field definition are still missing. This paper fills the gap by offering a thorough review of C-XAI approaches. We define and identify different concepts and explanation types. We provide a taxonomy identifying nine categories and propose guidelines for selecting a suitable category based on the development context. Additionally, we report common evaluation strategies including metrics, human evaluations and dataset employed, aiming to assist the development of future methods. We believe this survey will serve researchers, practitioners, and domain experts in comprehending and advancing this innovative field.

Figures (12)

• Figure 1: Word cloud generated from the titles, abstracts, and keywords of the reviewed C-XAI papers. The trend towards human understanding and interpretability is noticeable, in contrast to standard XAI approaches.
• Figure 2: Concepts and explanations provided by C-XAI methods and models. Some concepts require additional knowledge to be implemented.
• Figure 3: Categorization of the C-XAI methods with guidelines for selecting a suitable approach.
• Figure 4: Post-hoc Supervised Methods providing explanations in terms of class-concept relations. These approaches provide a set of samples annotated with concepts to the explained network to determine their influence on the output class.
• Figure 5: Post-hoc supervised methods providing an explanation in terms of Node-Concept Associations. This class of approaches associates concepts with internal nodes or filters of the network to increase the transparency of the network decision-making process.