[Paper Review] Explainable Artificial Intelligence Approaches: A Survey

Sheikh Rabiul Islam, William Eberle+2 more|arXiv (Cornell University)|Jan 23, 2021

Explainable Artificial Intelligence (XAI)Computer Science54 references88 citations

TL;DR

This survey reviews popular XAI methods using a credit default prediction case study, analyzes strengths/limits, and discusses quantifying explainability and human-centered AI paths.

ABSTRACT

The lack of explainability of a decision from an Artificial Intelligence (AI) based "black box" system/model, despite its superiority in many real-world applications, is a key stumbling block for adopting AI in many high stakes applications of different domain or industry. While many popular Explainable Artificial Intelligence (XAI) methods or approaches are available to facilitate a human-friendly explanation of the decision, each has its own merits and demerits, with a plethora of open challenges. We demonstrate popular XAI methods with a mutual case study/task (i.e., credit default prediction), analyze for competitive advantages from multiple perspectives (e.g., local, global), provide meaningful insight on quantifying explainability, and recommend paths towards responsible or human-centered AI using XAI as a medium. Practitioners can use this work as a catalog to understand, compare, and correlate competitive advantages of popular XAI methods. In addition, this survey elicits future research directions towards responsible or human-centric AI systems, which is crucial to adopt AI in high stakes applications.

Motivation & Objective

• Demonstrate popular XAI methods using a mutual loan default prediction task
• Compare the pros, cons, and competitive advantages of XAI methods from multiple perspectives
• Provide insights on quantifying explainability and evaluating human-centered AI under XAI
• Recommend directions toward responsible AI using explainability as a medium

Proposed method

• Describe and categorize XAI methods into intrinsically interpretable, model-agnostic, and example-based explanations
• Use a mutual test case (credit default prediction) with the Freddie Mac dataset to illustrate method workings
• Analyze methods from global vs local, ante-hoc vs post-hoc, and inherent vs emulated explainability perspectives
• Apply R's iml package to generate and compare explanations for the case study
• Consult and synthesize findings from recent surveys to augment analysis
• Present a framework for comparing explanations and discuss open questions and future research directions

Experimental results

Research questions

• RQ1 What are the competitive advantages and limitations of popular XAI methods across different explanations (local/global, ante/post-hoc, inherent/ emulated)?
• RQ2 How can explainability be quantified and evaluated, including human-in-the-loop assessments and proxy metrics?
• RQ3 How do XAI methods perform and communicate in a real-world high-stakes task like credit default prediction?
• RQ4 What directions are promising for responsible or human-centered AI when using XAI as a medium?

Key findings

• XAI methods offer a range of trade-offs in interpretability, fidelity, and applicability across models and tasks
• Global methods (e.g., PDP, ALE) provide overall model behavior while local methods (e.g., LIME, SHAP) explain instance-level predictions
• Feature importance, interaction analyses, and surrogate models help but can be affected by feature correlation and interactions
• Counterfactuals, prototypes, and example-based explanations provide human-friendly insights but face challenges like Rashomon effects and selection of representative instances
• There is a identified need for method-agnostic, local, and actual (not illusionary) explanations to support responsible, human-centered AI