Feature Weighted Rival Penalized EM for Gaussian Mixture Clustering: Automatic Feature and Model Selections in a Single Paradigm

The rival penalized expectation-maximization (RPEM) algorithm has demonstrated its powerful capability to perform the model selection automatically in the context of mixture model. However, the performance may be degraded when irrelevant variables are included. To overcome this drawback, we adopt the concept of feature salience as the feature weight to measure the relevance to the clusters in the subspace, and integrate it into the RPEM algorithm. The proposed algorithm distinguishes the probably redundant features and estimates the number of clusters automatically and simultaneously in a single learning paradigm. Experiments conducted on both synthetic and benchmark real data set have shown the efficacy of the proposed algorithm