Research On Multi-label Learning Algorithms And Applications To Tag Recommendations

The rapid development of information technology has brought the information overload problem, which results that it’s hard for users to get useful information from massive data. Automatic classification is one of the effective means to solve the problem and it has been widely used in many areas. Traditional classification usually assumes that instances are associated with only one label. However, in some areas instances are usually associated with multiple labels such as text, images, video, and so on. In these areas, traditional classification algorithms are not useful. Therefore, the study of multi-label data becomes an important research topic. With the rapid development of tagging website, tag recommendation has become one of the hottest research topics. This paper focuses on multi-label learning. In addition, Based on the fact that the results of multi-label learning and tag recommendation are both collection of elements, this paper also research on applying multi-label learning algorithms to tag recommendation.Firstly, the basic concepts of multi-label learning and tag recommendation are elaborated in this paper. We summarize related algorithms about multi-label learning and tag recommendation separately. Then we analyze the advantages and disadvantages of them. Secondly, we propose a multi-label learning algorithm based on label relations. In this algorithm, the meta-level features are simplified. The simplified meta-level features are as informative about the relation of instance to label as meta-level features and they also reduce the dimension of input space. In addition, we propose to a method that combines meta-level features with label space to get label relations. In the prediction stage, the label set is predicted by combining meta-level features with label relations. Experimental results on multi-label datasets show that this algorithm utilizes label relation effectively and it’s more applicable to datasets with strong label relations. Finally, we changed some implementations in the above algorithm to adapt to tag recommendation, which include the presentation of tag relations and the production of the recommended tag. We use tag co-occurrence to evaluate tag relations. In the recommendation stage, we propose to use tag vector to present the history information about the user and the item. The recommended tags are produced by combining the tag vector with tag relations. Experimental results on tagging datasets show the developed algorithm utilizes the tag relations better than other tag recommendation algorithms and the recommended results are more accurate.