Visual Model Selection using Feature Importance Clusters in Fairness-Performance Similarity Optimized Space

TL;DR

The problem addressed is the difficulty of choosing among many fairness-aware models that trade off predictive performance and fairness. The authors introduce an interactive framework that learns a Mahalanobis distance over model feature-importances via weakly supervised metric learning (ITML), transforms the importances, and clusters models with $k$-means to reveal archetypes along the fairness–performance frontier. Key contributions include a concrete pipeline (pairwise similarity constraints, ITML embedding, and composite-validation for $k$) that yields five interpretable archetypes, demonstrated on the Adult and Bank Marketing datasets with FairTree and Fair Logistic Regression as fairness-aware learners. This approach reduces cognitive load for stakeholders by providing a compact, navigable map of model behaviors tied to feature usage and domain-relevant trade-offs, enabling principled model selection in high-stakes contexts.

Abstract

In the context of algorithmic decision-making, fair machine learning methods often yield multiple models that balance predictive fairness and performance in varying degrees. This diversity introduces a challenge for stakeholders who must select a model that aligns with their specific requirements and values. To address this, we propose an interactive framework that assists in navigating and interpreting the trade-offs across a portfolio of models. Our approach leverages weakly supervised metric learning to learn a Mahalanobis distance that reflects similarity in fairness and performance outcomes, effectively structuring the feature importance space of the models according to stakeholder-relevant criteria. We then apply clustering technique (k-means) to group models based on their transformed representations of feature importances, allowing users to explore clusters of models with similar predictive behaviors and fairness characteristics. This facilitates informed decision-making by helping users understand how models differ not only in their fairness-performance balance but also in the features that drive their predictions.

Figures (8)

• Figure 1: Side‐by‐side comparison of $k=5$ cluster assignments in fairness–performance space using (left) raw feature importances versus (right) Mahalanobis‐transformed importances. The learned metric reduces cluster overlap, making group boundaries much clearer.
• Figure 2: Cluster assignments in the Mahalanobis‐transformed feature‐importance space (optimal $k=5$), plotted in the fairness–performance plane. Distinct clusters highlight clear trade‐off archetypes.
• Figure 3: Per‐cluster boxplots of the top eight average SHAP feature importances. Each subfigure shows one cluster, with identical vertical scales to facilitate cross‐cluster comparison.
• Figure 4: Heatmap of pairwise distance change $\Delta d_{ij} = d_{ij}^{\mathrm{before}} - d_{ij}^{\mathrm{after}}$ for FTC on the Adult dataset, sorted by increasing $\theta$. Deep blues show where ITML strongly contracts distances among similar models.
• Figure 5: Single‐attribute FTC on Adult dataset: clustered model distributions when only (a) Age, (b) Gender, or (c) Race is treated as the protected attribute. Each uses its own optimal $k$ from composite validation.