A Novel Contrastive Clustering Algorithm Based on K-means

Benefitting from the simplicity and ease of implementation, K-means algorithm and its variants have been extensively studied and applied. However, the similarity measure manner limits the performance of previous K-means methods. Specifically, existing methods only take the similarity between the data point and itself cluster centroid into account while failing to consider the relationship with other centroids. In this study, we propose a method called Contrastive K-means (CKm). For training, to learn satisfactory cluster centroids of input data points for clustering, CKm simultaneously considers the similarity/dissimilarity between positive/negative pairs when optimizing the cluster centroids of data. Experimental results on seven prevalent datasets substantiate that CKm surpasses other K-means and variants in performance.