Evaluating explainable artificial intelligence (XAI) methods suffers from a lack of reliable procedures, especially the absence of ground‑truth explanations. Most existing works assess explanations by measuring their fidelity to a black‑box model’s predictions. Fidelity only indicates how well an explanation reproduces the model’s output, not whether it faithfully reflects the underlying decision logic. Consequently, different explanations can achieve similar fidelity scores while offering contradictory or misleading interpretations.
This paper introduces a synthetic ground‑truth framework for XAI evaluation. The key idea is to create synthetic datasets through controlled interventions, allowing the importance of each input component to be predefined. This yields ground‑truth explanations that are directly aligned with the model’s behavior. The framework is instantiated on three data domains—binary images, tabular data, and time‑series—covering visual, structured, and sequential scenarios.
We benchmark nine widely used XAI methods (e.g., Grad‑CAM, SHAP, LIME) across these domains. Experimental results reveal substantial limitations: some methods capture only local patterns, others behave inconsistently across data types, and a few produce completely erroneous feature attributions.
These findings demonstrate that fidelity‑based evaluation is insufficient for assessing explanation correctness, and that synthetic, intervention‑based benchmarks are essential for a trustworthy assessment of explanation quality.
Review: The synthetic ground‑truth framework offers a controllable and reproducible environment for XAI evaluation, exposing current methods’ causal shortcomings and guiding future improvements.