Perception - Aware Clustering

This paper develops a clustering algorithm using the rate-distortion-perception (RDP) framework. First, we address the problem of clustering $n$ points into $k$ clusters using a distortion metric in the rate-distortion (RD) context. Then, we extend the scope to the RDP framework by introducing a perception constraint to the clustering distortion objective to better preserve the source distribution. We propose an algorithm for the extended objective and prove its convergence. Numerical experiments with samples from probability distributions demonstrate that the proposed algorithm aligns with the RDP tradeoff characteristics. We also demonstrate through experiments on real-world datasets that the proposed algorithm provides better semantic clustering by improving the classification accuracy in next-generation communication networks.