Improved affinity propagation clustering algorithms: a PSO-based approach

Traditional clustering algorithms such as K-means need to input the number of clusters before the start of the algorithm. Affinity propagation (AP) clustering algorithm solves this problem by considering each data point as a prospective cluster head (exemplar) and finding a set of appropriate exemplars by message passing. However, the AP clustering algorithm requires two parameters: preference and damping factor. Providing the parameters in advance poses the same issue faced in the traditional clustering algorithm. Moreover, all data points are not equally relevant for becoming cluster heads. To overcome these problems, we propose two parameter-free particle swarm optimization-based algorithms, PSO-APver1 and PSO-APver2. Furthermore, we introduce a novel version of mutant PSO where two cluster validity indices are used to judge the quality of the clustering solution. In PSO-APver2, we consider the internal data distribution using the square wave function to determine the initial preference value of data points. We conducted experiments on 8 real-world datasets to show the efficacy of our proposed algorithms over classic algorithms and two AP-based algorithms. We conducted the Friedman test followed by post hoc analysis to exhibit the significance of our work.