Automatically Affinity Propagation Clustering using Particle Swarm

Affinity propagation (AP) is a clustering algorithm which has much better performance than traditional clustering approach such as K-means algorithm. AP can usually find a moderate clustering number, but "moderate" usually may not be the "optimal". If we have found the optimal clustering number of AP, to estimate the input "preferences" (p) and the effective corresponding "preferences" (p) interval from the data sets is hard. In this paper, we propose a new approach called Automatically Affinity Propagation Clustering (AAP). Our AAP method is absolutely "automatic". AAP represents the issue of finding the optimal AP clustering and the corresponding "preferences" (p) interval as an optimization problem of searching optimal solution of the input "preferences" (p). AAP searches the "preferences" (p) space using Particle Swarm Optimization (PSO) algorithm, and evaluates the particles' fitness using clustering validation indexes. In order to prevent particles from flying out of defined region, we used Boundary Checking (BC) rule to check the validity of particles' positions of PSO. According to lots of AAP's independent runs results, we can find AP's optimal clustering number and estimate the corresponding "preferences" (p) interval. One artificial data set and several real-world data sets are presented to illustrate the simplicity and effectiveness of AAP.