Multi-criteria Rank-based Aggregation for Explainable AI

Different explanation methods can contradict each other: Popular explainability tools like LIME and SHAP sometimes provide conflicting explanations for the same AI prediction, meaning users cannot always rely on a single explanation method to understand why an AI system made a decision.

Quality metrics for explanations involve trade-offs: The paper identifies that explanation quality can be measured across multiple dimensions (complexity, faithfulness, and stability), but improving performance on one metric may worsen performance on another—similar to how making an explanation simpler might make it less accurate.

Combining multiple explanation methods improves reliability: By aggregating explanations from different methods using a weighted approach that balances competing quality metrics, the researchers demonstrated they could produce more robust and consistent explanations than relying on any single explainer alone.

Ranking-based evaluation better reflects real-world use: The paper proposes new ways to evaluate explanations that focus on ranked feature importance (which features matter most) rather than just individual feature scores, making evaluation metrics more practical for how explanations are actually used in practice.

Specific aggregation techniques work best for this problem: Among various mathematical approaches tested, TOPSIS and WSUM algorithms proved most effective at combining multiple quality metrics, providing practitioners with concrete recommendations for implementing more transparent AI systems.