Accepted Paper at ECML PKDD 2023

Authors

Mauritius Klein, Collin Leiber, Christian Böhm

European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML PKDD 2023),
18–22 September 2023, Torino, Italy

Abstract

Clustering heterogeneous data is an ongoing challenge in the data mining community. The most prevalent clustering methods are designed to process datasets with numerical features only, but often datasets consist of mixed numerical and categorical features. This requires new approaches capable of handling both kinds of data types. Further, the most relevant cluster structures are often hidden in only a few features. Thus, another key challenge is to detect those specific features automatically and abandon features not relevant for clustering. This paper proposes the subspace mixed-type clustering algorithm k-SubMix, which tackles both challenges. Its cost function can handle both numerical and categorical features while simultaneously identifying those with the biggest impact for a high-quality clustering result. Unlike other subspace mixed-type clustering methods, k-SubMix preserves inter-cluster comparability, as it is the first mixed-type approach that defines a common subspace for all clusters. Extensive experiments show that k-SubMix outperforms competitive methods and reduces the data's complexity by a simultaneous dimensionality reduction.