Semi-supervised Distributed Clustering with Mahalanobis Distance Metric Learning

Semi-supervised clustering uses a small amount of supervised information to aid unsupervised learning. As one of the semi-supervised clustering methods, metric learning has been widely used to clustering the centralized data points. However, there are many distributed data points, which cannot be centralized for the various reasons. Based on MPCK-MEANS framework [1] , the method of distributed Mahalanobis distance learning, called Distributed MPCK-MEANS, is proposed. Similar to MPCK-MEANS, Distributed MPCK-MEANS adopts EM iteration framework to performance clustering and metric learning simultaneously. To decrease the communication costs and protect the data security and privacy, corresponding transmission parameters are designed; furthermore, the transmission techniques are realized too. The proposed method is different to Parallel K-means, but is similar to Parallel K-means with some degrees. The proposed method is tested on artificial data sets and real data sets. Compared with Parallel K-means, Distributed MPCK-MEANS can effectively improve the quality of clustering.