A Kernel Enabled RPCL Algorithm

This paper presents a novel Kernel enabled Rival Penalised Competitive Learning (KRPCL) algorithm for clustering. Not only is it able to perform correct clustering without restriction on cluster shape, but also can automatically find the number of clusters. This algorithm generalizes the Rival Penalised Competitive Learning (RPCL)algorithm [15] by using the state-of-the-art kernel trick [2]. Moreover, under the framework of KRPCL, better seeds initialization can be achieved by using spectral analysis on the kernels used in KRPCL. As a third contribution, we show the RPCCL [16], proposed for handling the de-learning rate problem arising in RPCL, can also be generically kernel-enabled, which in turn motivates us to find more ways for specifying the delearning rate in KRPCL. The experimental results show superiority of KRPCL in both synthetical datasets and various real data collected from the UCI repository and the Internet newsgroups.