A New Similarity Measure Based Affinity Propagation for Data Clustering

摘要

A new unsupervised clustering technique called Affinity Propagation (AP) has been applied in many fields such as computer vision and bioinformatics. AP shows an improved performance than diverse reputable traditional clustering methods such as k -medoids, k -means. Normallyordinary clustering methods use a predefined set of cluster centers. However, all points in dataset considered as probable cluster center in AP. Potential centers emerged based on transmitting message among similarity pairs of data. The input similarity matrix of AP is calculated from thenegative Euclidean distance method, which is used to find the similarity between data points. Despite the good performance of AP, the negative Euclidean distance tends to be sensitive to deformation. In this paper, the drawback is addressed by proposing two similarity measure methods for measuringthe similarities between pairs of data points of a dataset. The proposed method is evaluated using Adjusted Rand index, Jaccard Coefficient and Fowlkes and Mallows. The experimental results show the proposed methods outperformed existing clustering method in accuracy.