Faithfulness Without Anchors: A Position Paper on the Structural Misalignments of Explainable AI Faithfulness Metrics for Feature Importance

Structural Misalignments in Faithfulness Metrics for Explainable AI

The field of Explainable Artificial Intelligence (XAI) has long aimed to enhance trust in machine learning models by making their decision-making processes more transparent to humans. However, a growing body of evidence suggests that current XAI methods often fall short of this goal. Instead of fostering trust, these tools may inadvertently promote overreliance on incorrect predictions, leading to algorithmic aversion. This misalignment between the model and its explanations is a critical issue, often referred to as unfaithfulness or infidelity.

Faithfulness, defined as the degree of approximation between an explanation and the model itself, is a key component of trustworthiness in XAI. However, existing metrics for quantifying faithfulness are limited in both feasibility and computational efficiency. These limitations hinder the ability to rigorously assess the reliability of explanations, which is essential for ensuring that XAI tools serve their intended purpose.

The paper identifies five major shortcomings in current faithfulness metrics. These include the inability to capture the complexity of model behavior, the lack of alignment with human interpretability, the computational expense of accurate quantification, the overemphasis on local explanations at the expense of global understanding, and the absence of standardized benchmarks for evaluation. These issues collectively contribute to the structural misalignments that undermine the effectiveness of XAI methods.

Addressing these challenges requires a reevaluation of how faithfulness is measured and interpreted. The authors propose four research avenues to improve the reliability of faithfulness metrics. These include the development of more robust and scalable evaluation frameworks, the integration of human-centric interpretability criteria, the exploration of hybrid approaches that combine local and global explanations, and the establishment of standardized benchmarks for comparison.

While the paper does not provide hands-on testing or specific benchmarks, it offers a critical perspective on the current state of XAI and highlights the need for more rigorous and aligned methods of quantifying faithfulness. This analysis underscores the importance of aligning XAI tools with the actual needs of users and the broader goals of trust and transparency in machine learning.

Source: arXiv AI. Analysis based on publicly available information; no hands-on testing is claimed.