Explainable cluster analysis: a bagging approach

A team of researchers has published a paper titled "Explainable cluster analysis: a bagging approach" on arXiv, proposing a novel framework to solve a major limitation in machine learning: the lack of explainability in clustering algorithms. Traditional methods like k-means group data but rarely reveal which specific features (e.g., customer age, purchase frequency) are responsible for forming those groups. The new method, developed by Federico Maria Quetti, Elena Ballante, Silvia Figini, and Paolo Giudici, directly addresses this black-box problem by adapting techniques from supervised learning.

The core innovation is an ensemble-based framework that integrates bagging (bootstrap aggregating) and random feature dropout. By creating multiple bootstrap resamples of the data and running clustering on each, the method aggregates the results into a more stable and robust consensus partition. Crucially, it calculates feature importance scores by measuring the mutual information between each feature and the estimated cluster labels across all iterations, weighted by the quality of each partition. This process, analogous to how Random Forests rank feature importance for prediction tasks, finally provides data scientists with a clear, quantifiable answer to the question: "Why are these observations grouped together?" The paper demonstrates the method's effectiveness on both simulated and real-world datasets, marking a significant step toward interpretable unsupervised learning.