A New Feature Weighted Affinity Propagation Clustering Algorithm

Data clustering is an effective method for data analysis and pattern recognition which has been applied in many fields such as image segmentation, machine learning and data mining. It is the process of splitting the multidimensional data into several groupings or clusters based on some similarity measures. A cluster is usually defined by a cluster center. Generally, the information of the features may differ from each other and the contributions to the clustering are different. The most meaningful features play an important role in explaining the differences among the samples, thus should be pay a more attention in the clustering process to get a exact grouping. In order to reflect the particular contributions of the features, this paper proposed a new features weighted affinity propagation clustering (AP) algorithm. In this method, all the features are evaluated by a feature analysis method. The training samples are those which are near to the centers in each group according to the affinity propagation cluster result. The similarity matrix of AP is updated on the weighted features and the new cluster result is obtained. The radius by which the training samples are determined is a very important parameter in our method. We study it by means of the sum of weighted distance between any two samples clustered in the same group corresponding to the cluster result. In order to demonstrate our method, three public data sets from UCI were used. The experiment results on the three dataset showed the superiority of the features weighted AP method.