Host: The Japan Society for Management Information
Recently, it has been reported that the performance of methods using the kernel learning is quite effective for document classification problems. Relevance Vector Machine (RVM) is one of the typical techniques in such methods. There is, however, a room to develop the performance for multivalued classification problems, although RVM is known as an excellent two-class classifier. Then, we propose in this paper a technique for multivalued classification by the redundant configuration of two-class classifiers and calculation of the posterior probability by using the property that RVM is a probabilistic model. The proposed technique is applied to a classification problem of newspaper articles to show its effectiveness, then, it is verified that the classification accuracy is improved.
