Predictive K-means with local models

The paper addresses a key tension in responsible AI: supervised classifiers often make accurate predictions but lack interpretability, while clustering provides understandable groupings without guaranteeing predictive usefulness. The proposed "predictive clustering" approach aims to achieve both goals simultaneously.

The authors introduce algorithms that transform data representation based on class densities before clustering, enabling k-means to discover clusters that are "as pure as possible with regards to the class labels"—meaning clusters that align with actual categories while remaining interpretable to humans.

Testing across multiple datasets, the new algorithms achieve prediction performance competitive with traditional supervised classifiers while maintaining the interpretability advantage of clustering, suggesting a practical path toward more explainable AI systems without sacrificing accuracy.

This work is relevant to responsible AI because it demonstrates that interpretability and performance need not be traded off against each other; algorithmic design choices can be made to optimize for both explainability and predictive power simultaneously.

The approach is particularly valuable for high-stakes applications where decision-makers need to understand why a system made a prediction (which clusters a person belongs to) rather than treating the model as a "black box," supporting more trustworthy and accountable AI deployment.