Automatic Clustering Using an Improved Particle Swarm Optimization

Abstract

Unsupervised data clustering is an important analysis in data mining. Many clustering algorithms have been proposed, yet most of them require predefined number of clusters. Unfortunately, unavailable information regarding number of clusters is commonly happened in real-world problems. Thus, this paper intends to overcome this problem by proposing an algorithm for automatic clustering. The proposed algorithm is developed based on a population-based heuristic method named particle swarm optimization (PSO). It overcomes two main issues in automatic clustering, namely determining number of clusters and cluster centroid. In the automatic clustering using PSO (ACPSO), the exploration is conducted by particles comprising of two sections. Herein, time-varying tuning parameter is applied. Furthermore, sigmoid function is employed to handle infeasible solution. In addition, K-means is applied to adjust the cluster centroids. Method validation using four benchmark datasets reveals that TPSO outperforms other two previous methods namely DCPSO, DCPG, and DCGA. Overall, ACPSO has better accuracy and consistency. 

A comprehensive survey of traditional, merge-split and evolutionary approaches proposed for determination of cluster number

“…Kao and Lee [162] introduced a combination of CPSO (Jarboui et al 2007) and K-means (KCPSO), where each particle tends to use K-means to reach better clustering result. Kuo and Zulvia [163] introduced a PSO based automatic clustering approach based on continuous activation-based encoding and the VI index. The proposed approach was compared with DCPSO, DCPG and DCGA on the Iris, Wine, Glass and Aggregation benchmark datasets.…”

Combining K-Means and K-Harmonic with Fish School Search Algorithm for data clustering task on graphics processing units

“…Soft computing methods have been used to bear down the limitations of the partitional clustering methods, such as K-Means. In the last decade, nature inspired optimization techniques were employed for Data Clustering, such as Simulated Annealing (SA) [6], Genetic Algorithms (GA) [7], Tabu Search (TS) [9], Differential Evolution (DE) [14], Gravitational Search Algorithm (GSA) [16,19], Black Hole (BH) algorithm [25], Intelligent Water Drop (IWD) [27], and mostly, swarm based algorithms, as Particle Swarm Optimization [31][32][33][34][35][36][37][38][39][40][41][42][43][44][45][46][47][48][49][50], Ant Colony Optimization (ACO) [4,8], HBMO (Honey Bee Mating Optimization) [5], ABC (Artificial Bee Colony) [10,24], Glowworm Swarm Optimization (GSO) [12], Firefly Algorithm (FA) [15,17,18], Bat Algorithm (BA) [20], Cat Swarm Optimization (CSO) [21], Wolf Search Algorithm (WSA) [23] and Cuckoo Search (CS) algorithm [11,26,28,…”

“…Numerous studies have been performed concerning clustering using SI . Most of them employ Particle Swarm Optimization (PSO) [31][32][33][34][35][36][37][38][39][40][41][42][43][44][45][46][47][48][49][50], the simplest SI algorithm, which is the SI paradigm that has received widespread attention in research.…”

Study and analysis of particle swarm optimization for improving partition clustering

“…Kuo et al [52] proposed an algorithm based on a population based search method known as PSO. It is mainly applied to overcome the issues in automatic clustering, which is determining the number of clusters in advance.…”