CHILLI: A data context-aware perturbation method for XAI

TL;DR

CHILLI tackles the problem of faithfulness in XAI by incorporating data context into local perturbation based explanations. It introduces context aware proximity and SMOTE inspired, domain representative perturbations to generate perturbations that reflect training data and local structure, and evaluates them against LIME on WebTRIS and MIDAS. The results show CHILLI yields substantially more faithful explanations (lower proxy-model error) and more contextually meaningful feature contributions, demonstrating that accounting for feature dependencies and data bounds improves trust in explanations. This approach enhances the practical utility of post hoc XAI in high risk domains by producing explanations that better reflect the base model behavior and real-world data constraints.

Abstract

The trustworthiness of Machine Learning (ML) models can be difficult to assess, but is critical in high-risk or ethically sensitive applications. Many models are treated as a `black-box' where the reasoning or criteria for a final decision is opaque to the user. To address this, some existing Explainable AI (XAI) approaches approximate model behaviour using perturbed data. However, such methods have been criticised for ignoring feature dependencies, with explanations being based on potentially unrealistic data. We propose a novel framework, CHILLI, for incorporating data context into XAI by generating contextually aware perturbations, which are faithful to the training data of the base model being explained. This is shown to improve both the soundness and accuracy of the explanations.

Figures (9)

• Figure 1: (a) The global decision space, with red and green representing decision regions for two classes. (b) The local area around an instance, with a set of perturbations shown as orange crosses.
• Figure 2: A comparison of the effect on proximity between two points of fixed distance, $d$, as the locality parameter, $\sigma$, is varied.
• Figure 3: Normalised WebTRIS training data shown in each feature dimension against the target 'Total Volume' feature. The 25 instances selected uniformly at random for evaluation are shown as the dark blue points.
• Figure 4: Normalised MIDAS training data shown in each feature dimension against the target 'Keswick Air Temperature' feature. The 25 instances selected uniformly at random for evaluation are shown as the dark blue points.
• Figure 5: (a) Perturbations (orange points) generated using LIME and CHILLI for a WebTRIS data instance, $x$. The predicted 'Total Volume' for each perturbation from the base model, $f(z)$, is shown on the vertical axis. Opacity of perturbations represent proximity to $x$. (b) Explanations produced by CHILLI and LIME, showing the feature coefficients of the linear proxy model fit to the perturbations in (a) representing the contribution of each feature towards the predicted target value, $f(x)$.