Does Dataset Complexity Matters for Model Explainers?

Different XAI tools produce inconsistent explanations: The study tested six major explainability tools (SHAP, DALEX, Eli5, LOFO, Ciu, and Skater) across 82 models and found that these tools don't always agree on which features are most important—meaning the "explanation" you get depends significantly on which tool you use, raising questions about reliability.

Dataset complexity significantly affects explainability: The research demonstrates that how complex a dataset is matters just as much as the explainability method itself. Simpler datasets may produce clearer, more consistent explanations across different tools, while complex datasets could yield conflicting results—an important consideration when deploying AI systems in real-world scenarios.

No single "best" explainability measure exists: The study found no consensus on which XAI measure generates the most reliable overall explainability ranking, suggesting that organizations cannot rely on a one-size-fits-all approach to understanding their AI models' decisions.

Explainability tools need standardization: With multiple competing tools producing different attribute importance rankings, there's a need for better standards and benchmarks in the XAI field to ensure consistent, trustworthy explanations—critical for responsible AI deployment where stakeholders need reliable insights into model behavior.