Mutually beneficial learning with application to on-line news classification

There are three common challenges in real-world classification applications, i.e. how to use domain knowledge, how to resist noisy samples and how to use unlabeled data. To address these problems, a novel classification framework called Mutually Beneficial Learning (MBL) is proposed in this paper. MBL integrates two learning steps together. In the first step, the underlying local structures of feature space are discovered through a learning process. The result provides necessary capability to resist noisy samples and prepare better input for the second step where a consecutive classification process is further applied to the result. These two steps are iteratively performed until a stop condition is met. Different from traditional classifiers, the output of MBL consists of two components: a common classifier and a set of rules corresponding to local structures. In application, a test sample is first matched with the discovered rules. If a matched rule is found, the label of the rule is assigned to the sample; otherwise, the common classifier will be utilized to classify the sample. We applied the MBL to online news classification, and our experimental results showed that MBL is significantly better than Naïve Bayes and SVM, even when the data is noisy or partially labeled. © 2007 ACM.