CohEx: A Generalized Framework for Cohort Explanation

TL;DR

CohEx introduces cohort explanations as a principled middle ground between global and local XAI, defining a generalized, data-driven framework based on supervised clustering to produce context-specific explanations. It iteratively recomputes local importances within cohorts and refines cohort definitions through supervised clustering, balancing generalizability and conciseness while enforcing locality and stability. The approach is evaluated on synthetic, bike-sharing, and MNIST tasks, showing improved locality, generalizability, and interpretable cohort structure compared to baselines like VINE, REPID, GALE, and hierarchical schemes. The work provides a concrete objective and metrics for cohort explanations and highlights practical considerations and future directions for robust, scalable cohort-level explainability in real-world settings.

Abstract

eXplainable Artificial Intelligence (XAI) has garnered significant attention for enhancing transparency and trust in machine learning models. However, the scopes of most existing explanation techniques focus either on offering a holistic view of the explainee model (global explanation) or on individual instances (local explanation), while the middle ground, i.e., cohort-based explanation, is less explored. Cohort explanations offer insights into the explainee's behavior on a specific group or cohort of instances, enabling a deeper understanding of model decisions within a defined context. In this paper, we discuss the unique challenges and opportunities associated with measuring cohort explanations, define their desired properties, and create a generalized framework for generating cohort explanations based on supervised clustering.

Figures (8)

• Figure 1: A motivating example: a synthesized patient binary classification problem.
• Figure 2: Two 1-D regression models with the same behavior on a specific cohort. The horizontal axis represents the feature, and the vertical axis represents the model output. The orange circle denotes the cohort.
• Figure 3: Running the six cohort explanation algorithms on the motivational example in Fig. \ref{['fig:motivation']}. Each color represents a different cohort. The labels in the figures denotes cohort explanation $(e_\mathrm{age}, e_\mathrm{family})$ of the cohort of the corresponding color.
• Figure 4: Running CohEx on the bike sharing dataset. Each column represents a cohort. (a-d) Cohort importance of the top five features in each cohort; (e-h) the top two most homogeneous categorical features in each cohort, measured in homogeneity; (i-l) the top two most abnormal continuous features, measured in the centroid difference w.r.t. dataset mean in standard deviations.
• Figure 5: Generalizabilty loss vs. number of cohorts for the four cohort explanation algorithms on the patient classification dataset. The solid lines represent the average clustering loss among ten trials, and the shaded areas are within one standard deviation. Note that REPID is only evaluated at 4, 9 and 16 cohorts, with a respective max tree depth of 2, 3 and 4.