[Paper Review] Explainable Artificial Intelligence: a Systematic Review

Giulia Vilone, Luca Longo|arXiv (Cornell University)|May 29, 2020

Explainable Artificial Intelligence (XAI)307 references90 citations

TL;DR

A systematic review that classifies XAI literature into four main clusters—reviews, notions, methods, and evaluation—and outlines state-of-the-art and future directions.

ABSTRACT

Explainable Artificial Intelligence (XAI) has experienced a significant growth over the last few years. This is due to the widespread application of machine learning, particularly deep learning, that has led to the development of highly accurate models but lack explainability and interpretability. A plethora of methods to tackle this problem have been proposed, developed and tested. This systematic review contributes to the body of knowledge by clustering these methods with a hierarchical classification system with four main clusters: review articles, theories and notions, methods and their evaluation. It also summarises the state-of-the-art in XAI and recommends future research directions.

Motivation & Objective

• Organize the vast XAI literature by defining clear boundaries and groupings.
• Present a hierarchical classification system for XAI research.
• Summarize state-of-the-art concepts, notions, and evaluation approaches in XAI.
• Recommend future research directions and open challenges in XAI.

Proposed method

• Search for explainability literature using Google Scholar with terms: 'explainable artificial intelligence', 'explainable machine learning', 'interpretable machine learning'.
• Perform two-phase article selection: phase 1 identifies ~200 peer-reviewed publications; phase 2 analyzes bibliographies to converge on ~100 articles.
• Classify selected works into four main categories: reviews, notions, methods, and evaluation.
• Construct a tree-like hierarchical map of the XAI literature with leaves representing individual articles.
• Synthesize notions related to explainability and outline evaluation notions and metrics from HCI literature.

Experimental results

Research questions

• RQ1 What are the main categories and boundaries of the explainable AI literature?
• RQ2 How are notions of explainability defined and operationalized across studies?
• RQ3 What methods exist for explainability, and how are they evaluated?
• RQ4 What are the identified future directions and challenges in XAI?
• RQ5 How do reviews organize and synthesize diverse XAI approaches across domains?

Key findings

• Four main categories of XAI literature were identified: reviews, notions, new methods for explainability, and evaluation of explainability.
• A tree-like hierarchical map shows the distribution and dependencies among categories, highlighting how reviews rely on notions, methods, and evaluation.
• Reviews cluster around application fields, construction approaches, theories and concepts, output formats, problem types, and generic/systematic reviews.
• Explanations can take textual, visual, or rule-based formats, and their effectiveness is tied to notions such as trust, causality, completeness, and intelligibility.
• Ethical and legal context (e.g., GDPR right to explanation) motivates the need for transparent and explainable systems.
• The literature emphasizes human-centric factors and interdisciplinary insights from HCI, cognitive science, and philosophy in defining explainability and assessing explanations.