Multi-Label Classification By Exploiting Relationship Of Labels

Traditional supervised learning works under the single label scenario, each example is associated with one single label. However, in multi-label classification, each example in a training set can be associated with multiple labels and the task of the multi-label classification which is different from the task of the traditional classification problem is to predict the proper label set for the unseen example. Nowadays, the key to successful multi-label learning is how to effectively exploit correlations between different labels to facilitate the learning process. However, many existing multi-label classification methods assume that the different labels are independent of each other and neglect the corrections between different labels; therefore they lost a lot of important information which is available from the training set and the performance of classifiers is badly affected.This paper presents a new method (multi-label classification by exploiting relationship of labels, MCER) to improve the performance of classifiers via exploring the correlations between different labels. MCER has two major aspects:(1) MCER adds a virtual label which will be used during prediction to the original label set and constructs a binary classifier for every pairwise label.(2) When constructing a binary classifier for a pairwise label, MCER gets a label subset from original label set according to mutual information and adds the subset to original feature set. In classification process, all classifiers give their results and the final result can be got according to the votes of each label. Extensive experiments on a variety of multi-label datasets show that our method can achieve better performance than the other multi-label methods.This research can be used in many domains, such as the categorization of textual data, functional genomics, semantic annotation of images and video and so on. It is very important for the development for those domains.