Dimensionality reduction in unsupervised learning of conditional Gaussian networks

This paper introduces a novel enhancement for unsupervised learning of cond itional Gaussian networks that benefits from feature selection. Our proposa l is based on the assumption that, in the absence of labels reflecting the cluster membership of each case of the database, those features that exhibi t low correlation with the rest of the features can be considered irrelevan t for the learning process. Thus, we suggest performing this process using only the relevant features. Then, every irrelevant feature is added to the learned model to obtain an explanatory model for the original database whic h is our primary goal. A simple and, thus, efficient measure to assess the relevance of the features for the learning process is presented. Additional ly, the form of this measure allows us to calculate a relevance threshold t o automatically identify the relevant features. The experimental results re ported for synthetic and real-world databases show the ability of our propo sal to distinguish between relevant and irrelevant features and to accelera te learning; however, still obtaining good explanatory models for the origi nal database.