Research On MIML Learning Method Of Weak Markup

MIML (Multi-Instance Multi-Label learning) is a new machine learning framework which is more suitable for representing complicated objects with multiple concepts. In recent years, researches on MIML have attracted much attention and many MIML algorithms have been developed. This framework has been found useful in diverse tasks especially those involving complicated data objects, such as image annotation, video annotation, web categorization, bioinformatics, etc. Previous studies typically assume that for every training example, all positive labels are tagged whereas the untagged labels are all negative. In many real applications such as image annotation, however, the learning problem often suffers from weak label; that is, users usually tag only a part of positive labels, and the untagged labels are not necessarily negative. To the best of our knowledge, the Multi-Instance Multi-Label learning with weak label setting has never been touched.In detail, main works of this thesis can be summarized as follows:(1) A Multi-Instance Multi-Label learning method MIMLwel is proposed which can solve weak label problem effectively. This method works by assuming that highly relevant labels share some common instances, and the underlying class means of bags for each label are with a large margin. This method extends an efficient block coordinate descend algorithm to solve this weak label problem. Experiments validate the effectiveness of MIMLwel in handling the weak label problem.(2) A Multi-Instance Multi-Label learning method MIMLwel-ins is proposed which can improve the performance on weak label problem by deleting the noise instance in each bag. This method works by assuming that the underlying class means of bags for each label will be with a larger margin after deleting the noise instance in each bag. In this way, this method can complete the missing labels with deleting the noise instance simultaneously. Experiments show that this problem deserves further exploration.