A Research Of Multi-Label Learning Method Based On Deep Forest

With the rapid development of the Internet,we have entered an epoch of informa-tion explosion,and machine learning has been widely studied and applied.Multi-label learning is an important research area of machine learning and multi-label classification problems are very common in real life.In multi-label learning,each instance is asso-ciated with multiple labels,and the crucial task is how to leverage label correlations in building models.As the new technology in machine learning,the deep forest opens up a new way different from the deep neural network.A series of researches on how to apply deep forest to solve the multi-label classification problems are carried out in this paper and the innovations are as follows.Firstly,Multi-Label Deep Forest(MLDF)is proposed,which firstly introduces the deep forest to the multi-label learning area and provides a new idea to this area.MLDF relies on the cascade structure of deep forest to explore and exploit the correlations.Experimental results show that MLDF can indeed explore the correlations,achieve the best performance among compared methods and has strong flexibility.Secondly,a method based on MLDF named Measure-Aware Multi-Label Deep Forest(MA-MLDF)is proposed,which can optimize different measures on different demands.Two mechanisms:Measure-aware Feature Reuse and Measure-aware Layer Growth are included to enable MA-MLDF to adjust the model guided by the specific performance measure.Experiments demonstrate that the performance of MA-MLDF is better than the compared methods and the two proposed mechanisms are effective.Thirdly,a method based on MA-MLDF named Semi-Supervised Multi-Label Deep Forest(SS-MLDF)is proposed,which can use unlabeled data to assist the training in labeled data when some instances are unlabeled.In SS-MLDF,the tri-training method is used to label the unlabeled data,and the mixup technology is used to enrich data to improve the generalization ability.The experimental results show that SS-MLDF is better than the compared methods.