Local Best Particle Swarm Optimization for Partitioning Data Clustering

This paper proposes a new method for partitioning data clustering using PSO. The Proposed methods LPSOC designed for hard clusters. LPSOC alleviate some of the drawbacks of traditional algorithms and the state-of-the-art PSO clustering algorithm. Population-based algorithms such as PSO is less sensitive to initial condition than other algorithms such as K-means since search starts from multiple positions. The proposed algorithm LPSOC is less susceptible to local minima than K-means or even gbest version of PSO. In gbest PSO, all centroids are encoded in a single particle. Thus, the global best particle is a complete solution to the problem because its encoding contains the best position found for the centroids of all clusters. We used the local version of PSO in LPOSC. LPSOC uses a neighborhood of particles for optimizing the position of each cluster centroid. The whole swarm represents a solution to the clustering problem. This representation is far less computationally expensive than standard gbest version. The LPSOC is tested using six datasets from different domains to measure its performance fairly. LPOSC is compared with standard PSO for clustering and K-means. The results assure that the proposed method is very promising.