A novel features partition algorithm for semi-supervised categorization

This paper investigates the co-training algorithm and its assumption that the features set can be split into two compatible and independent views. However, the assumption is usually violated to some degree in real-world application, especially for independence. Sometimes the natural feature split does not exist. So a novel features partition algorithm, named Partition-MID, has been proposed. We give the formulas to estimate mutual independence between two features, feature and sub-view, sub-view and sub-view. And then the features with weaker mutual independence are classified in the same sub-view, whereas those with stronger mutual independence are in separate sub-views. Theoretical proof and experiments results both prove the proposed partition method can effectively split features set into two sub-views with higher independence. Based on Partition-MID, a new semi-supervised categorization algorithm, named SC-PMID, is developed. Utilizing unlabeled data together with labeled data, SC-PMID algorithm can significantly improve classification precision, especially when labeled data is sparse.