MASALA: Model-Agnostic Surrogate Explanations by Locality Adaptation

Existing explanation methods require difficult manual tuning: Popular XAI methods like LIME require users to manually set a "locality size" hyperparameter that determines which parts of the input space to examine, and the paper demonstrates this choice significantly impacts the quality and reliability of explanations generated.

One-size-fits-all explanations are inadequate: The research shows that using a single locality size to explain all predictions is fundamentally flawed—different predictions may require different-sized regions to accurately capture why a model made its decision, meaning explanations need to be customized per instance.

MASALA automatically adapts to each prediction: The proposed method eliminates manual hyperparameter tuning by automatically determining the appropriate local region for each individual prediction, making XAI more accessible and less dependent on user expertise.

Improved reliability of explanations: Experiments on real datasets (PHM08 and MIDAS) demonstrate that MASALA produces explanations with higher "fidelity" (accuracy in representing model behavior) and "consistency" (reliability across similar cases) compared to existing methods like LIME and CHILLI.

Practical benefit for responsible AI deployment: By removing the need for sensitive hyperparameter choices and producing more faithful explanations automatically, MASALA makes it easier for organizations to generate trustworthy, auditable explanations of AI decisions—a key requirement for responsible AI systems.