Deep Plug-and-Play Clustering with Unknown Number of Clusters

Published: 19 May 2023, Last Modified: 19 May 2023Accepted by TMLREveryoneRevisionsBibTeX
Abstract: Clustering is an essential task for the purpose that data points can be classified in an unsupervised manner. Most deep clustering algorithms are very effective when given the number of clusters K. However, when K is unknown, finding the appropriate K for these algorithms can be computationally expensive via model-selection criteria, and applying algorithms with an inaccurate K can hardly achieve the state-of-the-art performance. This paper proposes a plug-and-play clustering module to automatically adjust the number of clusters, which can be easily embedded into existing deep parametric clustering methods. By analyzing the goal of clustering, a split-and-merge framework is introduced to reduce the intra-class diversity and increase the inter-class difference, which leverages the entropy between different clusters. Specifically, given an initial clustering number, clusters can be split into sub-clusters or merged into super-clusters and converge to a stable number of K clusters at the end of training. Experiments on benchmark datasets demonstrate that the proposed method can achieve comparable performance with the state-of-the-art works without requiring the number of clusters.
Submission Length: Regular submission (no more than 12 pages of main content)
Changes Since Last Submission: First round revision: 1. Fixe typos and incorrect expressions 2. Complete assumptions and proof of convergence 3. Add sample complextiy analysis 4. Add distance metric analysis 5. Refine running time analysis Second round revision: Supplement lemma3, lemma4 and Propositon 5 and correspoing proofs to claim convergence. Camera ready version: 1. Further explain Proposition 5 to claim convergence. 2. Add proofs of Lemma 4 for better understanding of Proposition 5 3. Add complelet proofs of Proposition 1, Proposition 2 Lemma3 and Propositon 5 in the appendix. 4. Further clarify the effect of the introduced hyper-parameter $\lambda$ in ablation study
Assigned Action Editor: ~bo_han2
License: Creative Commons Attribution 4.0 International (CC BY 4.0)
Submission Number: 522