Research On Label Relationship Exploitation In Multi-Label Learning

In traditional machine learning research, each object is associated with only one label describing its semantic concept. However, in many real world applications, one object can be associated with multiple semantic concepts simultaneously. Multi-label learning is a framework to learn such objects, and has achieved great success in many applications. It is worth noticing that if we learn each label independently, the output space of multi-label learning will grow exponentially with the number of labels, and much more training examples will be needed to discriminate labels from each other,leading to high cost on both memory and time. Moveover, models on minority labels will suffer from poor performance due to the lack of training examples. Therefore,effective exploitation of label relationship is the core task of multi-label learning. This dissertation studies several important issues on label relationship exploitation in multi-label learning, and main results are summarised as follows.1. An effective approach MAHR is proposed, which does not require prior knowl-edge of the label relationship, and even outputs an estimation for the label relationship. Most existing methods require the label relationship be given as a prior knowledge, and are easy to suffer from overfitting when such knowledge is not available. In this dissertation, we propose a novel approach MAHR. By au-tomatically reusing models between labels, MAHR achieves strong generalization performance, and outputs an estimation for the label relationship. The effectiveness of MAHR is validated by both theoretical analysis and experimental results.2. An adaptive approach ML-LOC is proposed, which exploits the label relation-ship locally. Most existing methods assume that label relationship is helpful to all examples. However, in real applications, each label relationship may only work on some specific examples. By adaptively constructing a label relationship code vec-tor, the proposed approach ML-LOC can model the local influence of each label relationship. Experimental results show that ML-LOC has stronger generalization ability than approaches that globally exploit the label relationship.3. A fast approach MIMLfast is proposed, which exploits the label relationship for multi-instance multi-label learning. Most existing multi-instance multi-label methods can handle only small problems. By automatically mapping the com-plex original space to a shared low-dimensional subspace, the proposed approach MIMLfast can efficiently optimize the relative rank between labels. Theoretical analysis and experimental study validate the advantage of MIMLfast on both effec-tiveness and efficiency.4. Two active learning approaches AUDI and QUIRE are proposed, which ex-ploit the label relationship to reduce the number of human labeled examples.Most existing multi-label active learning methods ignore the label relationship, and consider only single factor to select query examples. By exploiting the label re-lationship with indirect and direct strategies, respectively, the proposed two ap-proaches consider two important factors: informativeness and representativeness for example selection. Experimental results show that AUDI and QUIRE signifi-cantly reduce the number of labeled examples.