A New Semi-Supervised Subspace Clustering Algorithm on Fitting Mixture Models

Abstract

We propose a new subspace clustering algorithm (EPSCMIX), which is based on the feature saliency measure that is obtained by using both the Emerging Patterns algorithm and the EM algorithm, for the analysis of microarray data. For the model selection, it employs a novel agglomerative step together with MDL criterion. And, we present the result of comparative experiments between AIC, MDL and minimum message length (MML) used to determine a criterion for our algorithm. The robustness of using emerging patterns based on mixture models, as well as using the Gaussian mixture model for subspace clustering, was demonstrated on both synthetic and real data sets. In experiments, it also certified that a new agglomerative method that merges mostly correlated components with MDL consistently worked better than the one that removes weak weight components.