Unified Explanations in Machine Learning Models: A Perturbation Approach

Explanation methods need consistency checks: The paper identifies "inconsistencies between XAI and modeling techniques" that can undermine trust in AI explanations. This suggests that before relying on an AI system's explanation for a decision, organizations should verify that the explanation method actually aligns with how the model behaves.

SHAP explanations may not be universally reliable: The authors conduct systematic testing of SHAP, a widely-used explanation technique, using perturbation analysis across multiple machine learning models. Their findings on "quantifiable similarity amongst explanation models" suggest that SHAP's explanations can vary depending on the specific model and dataset, meaning users shouldn't assume SHAP explanations are equally trustworthy in all contexts.

Feature importance rankings can change depending on conditions: The research examines "dynamic inference" scenarios and compares them to static cases, finding that the importance of features in explaining a model's decision may shift under different conditions. This is important for real-world applications where data or conditions change over time—explanations that were valid yesterday may not hold today.

A framework for evaluating explanations is needed: By proposing "a taxonomy for feature importance methodology" and metrics to measure explanation quality, the paper suggests that the AI field needs standardized ways to test whether explanations are actually reliable, rather than simply accepting them at face value.