Weakly Supervised Multi-label Learning

In most practical machine learning tasks,one instance is associated with multiple labels.while in the whole data set,only part of instances,or no instances are completely annotated.To learn from such kind of data is called Weakly Supervised Multi-Label(WSML)Learning.This problem has always been encountered in practical applications and becomes a new challenge in machine learning community.This paper tries to solve the weakly supervised multi-label learning problem,including the following aspects,1.Weakly supervised multi-label learning with fully annotated instances.When some of the instances have been fully annotated,this paper proposes a novel CUR matrix approximation algorithm,which can complete the matrix based on fully ob-served rows corresponding to fully annotated instances.Theoretical results show that the proposed algorithm is effective under practical conditions.Experiments validate the strength of the proposed method.2.Weakly supervised multi-label learning without fully annotated instances.When there is no fully annotated instance,this paper proposes a novel matrix completion technique,which can exploit side information,such as instances' features and la-bel correlation,to complete the matrix.Theoretical results show that the proposed algorithm is effective under practical conditions.Experiments validate the strength of the proposed method.3.Top-ranked weakly supervised multi-label learning.Labels ranked at the top are more important than labels ranked at the bottom.To fulfill this requirement,this paper proposes a novel criterion,PRO LOSS,concerning the prediction on all labels as well as the rankings of only relevant labels.We also propose correspond-ing algorithm to optimize PRO Loss.Experiments showed that our proposals are effective compared to baselines.4.Real-output weakly supervised multi-label learning.Classical multi-label learn-ing using a discrete 0/1 value to annotate while it is required to assign a real value to express the relativeness of a label to an instance.For this problem,we propose an objective based on trace norm minimization to exploit the correlations between la-bels.We develop a proximal gradient descent algorithm and an algorithm based on alternating direction method of multipliers.Experiments validate the effectiveness of our proposal.