MASALA: Model-Agnostic Surrogate Explanations by Locality Adaptation

TL;DR

MASALA addresses the locality selection problem in local XAI by automatically discovering instance-specific linear regions of base-model behaviour through a deterministic clustering of input-feature distributions and predictions. It fits a multivariate linear regression surrogate to points sharing the same linear region, enabling faithful and consistent explanations without a fixed locality hyperparameter. Across PHM08 and MIDAS, MASALA outperforms LIME and CHILLI in explanation fidelity and consistency, demonstrating the value of automatic locality adaptation for model-agnostic explanations. This work enables more trustworthy, reproducible explanations for high-stakes applications and offers a practical alternative to perturbation-based local surrogates.

Abstract

Existing local Explainable AI (XAI) methods, such as LIME, select a region of the input space in the vicinity of a given input instance, for which they approximate the behaviour of a model using a simpler and more interpretable surrogate model. The size of this region is often controlled by a user-defined locality hyperparameter. In this paper, we demonstrate the difficulties associated with defining a suitable locality size to capture impactful model behaviour, as well as the inadequacy of using a single locality size to explain all predictions. We propose a novel method, MASALA, for generating explanations, which automatically determines the appropriate local region of impactful model behaviour for each individual instance being explained. MASALA approximates the local behaviour used by a complex model to make a prediction by fitting a linear surrogate model to a set of points which experience similar model behaviour. These points are found by clustering the input space into regions of linear behavioural trends exhibited by the model. We compare the fidelity and consistency of explanations generated by our method with existing local XAI methods, namely LIME and CHILLI. Experiments on the PHM08 and MIDAS datasets show that our method produces more faithful and consistent explanations than existing methods, without the need to define any sensitive locality hyperparameters.

Figures (7)

• Figure 1: a) A surrogate fit within a locality which is too large leading to an inaccurate linear approximation of the non-linear decision boundary b) A surrogate fit within a locality which is too small and therefore, captures irregularities rather than the impactful linear trend. c) An explanation generated within an appropriate sized locality, which represents the true model behaviour in the immediate vicinity of the target instance. d) Appropriate explanations generated for three instances using localities of different sizes
• Figure 2: Distribution of 2 features from the same dataset against the model predictions and clustered into linear regions.
• Figure 3: Error of explanations generated using CHILLI with different kernel widths for random instances from PHM08.
• Figure 4: Explanations generated using LIME, CHILLI and MASALA, 10 times for the same instance from the MIDAS dataset. Each column is a single explanation with the colour of the square indicating the linear relationship each feature has towards the target variable. The kernel width setting used for LIME and CHILLI is shown in parentheses.
• Figure :