Explainable AI needs formalization

TL;DR

Current XAI methods lack formal problem definitions and evaluation criteria, limiting their reliability and utility in understanding and improving machine learning models.

Contribution

The paper advocates for formalizing XAI problems and developing targeted metrics to improve explanation correctness and validation.

Findings

Popular XAI methods attribute importance independently of prediction targets

Current XAI approaches are not evaluated against well-defined correctness criteria

Formal problem definitions can lead to better explanation validation

Abstract

The field of "explainable artificial intelligence" (XAI) seemingly addresses the desire that decisions of machine learning systems should be human-understandable. However, in its current state, XAI itself needs scrutiny. Popular methods cannot reliably answer relevant questions about ML models, their training data, or test inputs, because they systematically attribute importance to input features that are independent of the prediction target. This limits the utility of XAI for diagnosing and correcting data and models, for scientific discovery, and for identifying intervention targets. The fundamental reason for this is that current XAI methods do not address well-defined problems and are not evaluated against targeted criteria of explanation correctness. Researchers should formally define the problems they intend to solve and design methods accordingly. This will lead to diverse use-case-dependent notions of explanation correctness and objective metrics of explanation performance that can be used to validate XAI algorithms.