Affinity Propagation Clustering Using Centroid-Deviation-Distance Based Similarity

Abstract

Clustering is a fundamental and important task in data mining. Affinity propagation clustering (APC) has demonstrated its advantages and effectiveness in various domains. APC iteratively propagates information between affinity samples, updates the responsibility matrix and availability matrix, and employs these matrices to choose cluster centroid (or exemplar) of the respective clusters. However, since it chooses the sample points as the exemplars, these exemplars may not be the realistic centroids of the clusters they belong to. There may be some deviation between exemplars and the realistic cluster centroids. As a result, samples near the decision boundary may have a relatively large similarity with other exemplar they don’t belong to, and they are easy to be clustered incorrectly. To mitigate this problem, we propose an improved APC based on centroid-deviation-distance similarity (APC-CDD). APC-CDD firstly takes advantages of k-means on the whole samples to explore the more realistic centroid of the cluster, and then calculates the approximate centroid deviation distance of each cluster. After that, it adjusts the similarity between pairwise samples by subtracting the centroid deviation distance of the clusters they belong to. Next, it utilizes APC based on centroid-deviation-distance similarity to group samples. Our empirical study on synthetic and UCI datasets shows that the proposed APC-CDD has better performance than original APC and other related approaches.