Multi-label Learning With Label Enhancement

In traditional supervised learning,each training example is represented by a single instance while associated with a single class label.In real-world applications,each training instance may be associated with multiple class labels.The task of multi-label learning is to predict a set of relevant labels for an unseen instance.Traditional multi-label learning algorithms treat each class label as a logical indicator of whether the corresponding label is relevant or irrelevant to the instance,i.e.,+1 represents relevant to the instance and-1 represents irrelevant to the instance.Such label represented by-1or +1 is called logical label.Logical label cannot reflect different label importance.However,for real-world multi-label learning problems,the importance of each possible label is generally different.For real-world applications,it is difficult to obtain the label importance directly.Thus we need a method to reconstruct the essential label importance from the logical multi-label data.To solve this problem,this thesis assumes that each multi-label instance is described by a vector of latent real-valued labels,which can reflect the importance of the corresponding labels.Such label is called numerical label.The process of reconstructing numerical labels from logical multi-label data is called Label Enhancement.Some methods with similar function to label enhancement have been proposed in the past years.These label enhancement methods can be grouped into two types,i.e.,fuzzy label enhancement and graph-based label enhancement.Fuzzy label enhancement approaches constitute fuzzy membership degree for each label via fuzzy theory.Graph-based label enhancement approaches transform logical labels into numerical labels via utilizing the topological structure in the feature space.There have been some existing multi-label learning approaches based on label enhancement.These approaches are all two-stage approaches: the numerical labels are first reconstructed via label enhancement,and then the predictive model is trained according to the reconstructed labels.This thesis proposes a single-stage approach called LEMLL,i.e.,Label Enhanced Multi-Label Learning.LEMLL incorporates regression of the numerical labels and label enhancement into a unified framework,where numerical labels and predictive model are jointly learned.Extensive comparative studies validate that the performance of multi-label learning can be improved significantly with label enhancement and LEMLL can effectively reconstruct latent label importance from logical multi-label data.