Multi-Objective Clustering Based on Congestion Games With Player-Specific Cost Functions

Cluster analysis, the unsupervised identification of homogeneous groups within data, has become essential in numerous domains due to its broad applicability. As a central task in unsupervised learning, clustering aims to uncover the underlying structure of data by grouping similar instances without relying on predefined labels. This paper proposes a multi-objective clustering technique based on concepts from game theory. In particular, we formulate the clustering problem using Singleton Congestion Games with Player-Specific Cost Functions, a sub-class of non-cooperative simultaneous-move games, in which cluster heads act as players and data objects are treated as resources. This clustering model is solved using the Nash equilibrium as the solution concept, which can be computed in polynomial time, thereby ensuring favourable computational complexity for the proposed algorithm, MOCA-II. Experimental results on diverse synthetic and real-world datasets demonstrate that MOCA-II outperforms both mono-objective and multi-objective state-of-the-art clustering methods in terms of clustering quality and efficiency.