Supervised Clustered Interpretability: Explainable Subgroup Discovery via Cluster Separation and Partial Dependence Disparity

In this paper, we introduce the novel concept of clustered interpretability to enhance the transparency and explainability of artificial intelligence models in data sets with heterogeneous feature-outcome relationships. Traditional interpretability techniques often focus on global or local explanations, providing either an overview of the model’s behavior across the data set or explanations for individual predictions. However, these approaches can fall short when modeling clustered data sets, in which interpretable feature-outcome relationships across subgroups are desirable. Furthermore, existing research on interpretable clustering generally focuses on the interpretability of cluster assignment decisions in unsupervised settings rather than the interpretability of feature effects in supervised settings. To address this gap, our framework proposes a novel metric—the separation-disparity index—that prioritizes between-cluster variance while minimizing unnecessary heterogeneity in the clusters’ feature-outcome relationships. We use this metric alongside particle swarm optimization (PSO) to learn interpretable cluster solutions, incorporating novel modifications into the PSO algorithm to encourage high-quality solutions. We then demonstrate the value of our framework on synthetic data and a real-world healthcare data set of Parkinson’s disease patients.

History: This paper has been accepted by Kaushik Dutta for the Special Issue on Responsible AI and Data Science for Social Good.

Supplemental Material: The software that supports the findings of this study is available within the paper and its Supplemental Information ( https://pubsonline.informs.org/doi/suppl/10.1287/ijoc.2024.0657 ) as well as from the IJOC GitHub software repository ( https://github.com/INFORMSJoC/2024.0657 ). The complete IJOC Software and Data Repository is available at https://informsjoc.github.io/ .