Daniel Radrigan

Interpretability in classification models focuses on justifying the assignment of an element to a specific class, whereas in clustering the goal is to identify the relevant variables that characterize and differentiate groups. While K-Means centroids represent a classical method to describe clusters, this approach has limitations as it does not define clear boundaries nor highlight the truly representative features. In this study, we propose a post-clustering interpretation approach, in which, after applying the most suitable clustering algorithm, interpretable classifiers such as DSGD, RIPPER, or decision trees are employed to generate explicit rules that describe the structure of the groups. The contrast between centroids and rules reveals that only those features aligned with rules of high mass and low uncertainty are effectively representative. Our experiments, conducted on both synthetic and real datasets, show average accuracy levels above 90%, with variations depending on the type of data and the classifier used. While decision trees perform well on non-separable datasets albeit with high complexity, DSGD provides richer and more consistent interpretations in scenarios dominated by numerical variables, and RIPPER proves advantageous in contexts with diffuse clusters. These results demonstrate that combining clustering with interpretable classifiers constitutes an effective strategy to enhance the understanding of the underlying data structure.