关键词: 聚类, 次胜者受罚竞争学习算法, 样本密度, 聚类数目, 聚类中心

Abstract: The original Rival Penalized Competitive Learning (RPCL) algorithm ignores the influence of the geometry structure of a dataset on the weight variation of its nodes. A new RPCL algorithm proposed by Wei Limei et al. (WEI LIMEI, XIE WEIXIN. A new competitive learning algorithm for clustering analysis. Journal of Electronics, 2000, 22(1): 13-18) overcame the drawback of the original RPCL by introducing the density of samples to adjust the weights of nodes, while the density was not much objective. This paper defined a new density for a sample according to the pattern distribution of samples in a dataset, and introduced the density into the adjusting for the weights of nodes in RPCL to overcome the disadvantages of the available RPCL algorithms. The authors' improved RPCL algorithm was tested on some well-known datasets from UCI machine learning repository and on some synthetic data sets with noisy samples. The accuracy of determining the number of clusters of a dataset and the run time and the clustering error of the algorithms were compared. The Rand index, the Jaccard coefficient and the Adjust Rand index were used to analyze the performance of the algorithms. The experimental results show that the improved RPCL algorithm outperforms the original RPCL and the new RPCL proposed by WEI LIMEI et al. greatly, and achieves much better clustering results and has a stronger anti-interference performance for noisy data than that of the other two RPCL algorithms. All the analyses demonstrate that the improved RPCL algorithm can not only determine the right number of clusters for a dataset according to its sample distribution, but also uncover the suitable centers of clusters and advance the clustering accuracy as well as approximate the global optimal clustering result as fast as possible.

Key words: clustering, Rival Penalized Competitive Learning (RPCL) algorithm, sample density, cluster number, cluster center