Interpreting Categorical Data Classifiers using Explanation-based Locality

Local surrogate explanation methods are a popular class of post-hoc interpretability approaches that explain the rationale of machine learning models in the locality of every particular instance. Fidelity, which refers to the accuracy of explanation methods in imitating the actual behavior of a model, is highly affected by their strategy for identifying the locality of instances. To find the locality of an instance, we need to calculate the distance between the instance and perturbed data points concerning categorical and numerical features. While the distance of numerical features can be measured precisely, the existing works usually adopt a coarse-grained or imprecise approach for comparing categorical features. This is especially problematic in the categorical data setting, where defining a representative locality demands fine-grained semantic similarity information between categories. In this paper, we propose a locality generation approach for categorical data classifiers that makes no assumption about domain knowledge and infers categorical similarities by relying on the model's explanations. Further, we devise a multi-centered sampling approach based on the derived similarity information that, compared to the conventional instance-centered technique, captures the local behavior of the model more effectively. Moreover, we develop a knowledge-based locality generation approach based on knowledge graphs to benchmark our explanation-based method against a scenario where the similarity information is provided by a domain expert. The experiments conducted on various data sets demonstrate the efficacy of our approach in generating faithful explanations.