Performance Analysis of the Competitive Learning Algorithms on Gaussian Data in Automatic Cluster Selection

Abstract

A clustering problem occurs when an object needs to be assigned into a predefined group or class based on a number of observed attributes related to that object. The existing Competitive Learning (CL) algorithm and its variants (Frequency Sensitive Competitive Learning (FSCL), Rival Penalized Competitive Learning (RPCL), and Rival Penalized Controlled Competitive Learning (RPCCL)) have provided an appealing way to perform data clustering without knowing the exact number of clusters prior to clustering. This paper studies and analyzes the performance of these algorithms. The experimental results have been analyzed on the 2-D Gaussian data with the learning rate parameter kept same for all algorithms. The result showed that if number of output clusters is chosen equal to the number of clusters present in the input data then the performance for all the algorithms remains almost equal but when this number is chosen larger than the clusters present, then the RPCCL outperforms the other algorithms. Thus RPCCL gives the best performance in automatic cluster selection and we can use this feature of RPCCL algorithm in various useful applications like cluster analysis, curve detection, image segmentation, medical data analysis etc.