Performance Evaluation of Affinity Propagation Approaches on Data Clustering

Abstract

Classical techniques for clustering, such as k-means clustering, are very sensitive to the initial set of data centers, so it need to be rerun many times in order to obtain an optimal result. A relatively new clustering approach named Affinity Propagation (AP) has been devised to resolve these problems. Although AP seems to be very powerful it still has several issues that need to be improved. In this paper several improvement or development are discussed in , i.e. other four approaches: Adaptive Affinity Propagation, Partition Affinity Propagation, Soft Constraint Affinity propagation, and Fuzzy Statistic Affinity Propagation. and those approaches are be implemented and compared to look for the issues that AP really deal with and need to be improved. According to the testing results, Partition Affinity Propagation is the fastest one among four other approaches. On the other hand Adaptive Affinity Propagation is much more tolerant to errors, it can remove the oscillation when it occurs where the occupance of oscillation will bring the algorithm to fail to converge. Adaptive Affinity propagation is more stable than the other since it can deal with error which the other can not. And Fuzzy Statistic Affinity Propagation can produce smaller number of cluster compared to the other since it produces its own preferences using fuzzy iterative methods.