Reprodcibility of Contrastive Clustering

Anonymous

Keywords: clustering, contrastive learning

TL;DR: Reproduce the results from the orginal paper and improve the proposed model

Abstract: In the paper "Contrastive Clustering" (AAAI 2021), authors proposed an innovative contrastive learning algorithm in clustering, namely Contrastive Clustering (CC), which can remarkably outperform 17 traditional clustering methods including K-means, PICA in all evaluation metrics. Unlike traditional clustering, CC simultaneously conducts the instance-level and cluster-level contrastive learning in the row and column space of feature matrix, leading to an improvement of clustering performance. In order to verify the performance of their CC algorithm, we reproduce results from their paper. Specifically, we re-implement the results of Figure 3, Table 2, and Figure 4 from the paper to evaluate the reproducibility. Then We apply CC on other datasets to test the robustness of CC algorithm. In addition, we conduct ablation study to examine the importance of contrastive heads to the model performance. Finally, we successfully improve the performance of proposed CC model by modifying their loss function.

Paper Url: https://arxiv.org/abs/2009.09687