Co-training Based on Ensemble Learning.docVIP

Co-training Based on Ensemble Learning 　　（Computing school, Jingchu University of Science and Technology, Jingmen Hubei 448000, China） 　　Abstract. In many practical data mining applications, unlabeled training examples are readily available but labeled ones are fairly expensive to obtain. Semi-supervised learning algorithms such as co-training address this problem by using unlabeled data together with labeled data to improve the performance. Standard co-training has the assumption of two sufficient and independent views, which is not satisfied in many real-world application domains. In this paper, an ensemble of diverse classifiers is used for semi-supervised learning that requires neither two sufficient and independent views nor different base learning algorithms. The comparative studies with other semi-supervised learning methods demonstrate the effectiveness of the proposed approach. 　　Key words: machine learning; semi-supervised learning; co-training; classification; ensemble learning 　　1． Introduction 　　Machine learning and data mining has the ability to learn from data automatically and improve the decision-making process with experience. In many practical data mining applications such as web page classification, unlabeled training examples are readily available but labeled ones are fairly expensive to obtain because they require human effort. Therefore, semi-supervised learning that exploits unlabeled examples in addition to labeled ones has become a hot topic [1]. 　　Among various efforts of SSL, co-training is a well-known approach that was originally proposed in [2]. The standard co-training algorithms requires two sufficient and independent views, that is ,the attributes be naturally partitioned into two sets, each of which is sufficient for learning and conditionally independent to the other given the class label. 　　Unfortunately, the requirement of co-training for two sufficient and independent views can hardly be met in most scenarios [3, 4,