Noise Tolerated Weekly Supervised Multi-Label Learning With Label Enrichment

In multi-label learning tasks,one sample corresponding to multiple classes,which has wider application background than single label learning.However,it is difficult obtain complete labeling information.The learning of such data samples is weakly supervised multi-label learning.In addition to label missing problem,real world datasets often contain both feature noise and label noise,but most traditional multi-label learning algorithms are based on the fact that the training samples have no noise,or just consider one of them.Ignoring any kind of noise will affect the prediction performance of the multi-label learning algorithm,so how to build a noise-tolerant multi-label learning algorithm is also a great challenge.In response to the above challenges,this thesis focuses on the weakly supervised multi-label learning model algorithm,based on laebl enrichmegnt,two multi-label learning models are proposed by introducing hybrid noise model and multi view learning.The main contributions of this thesis including the following three aspects:1.For the single view scene with feature noise and label noise appear at the same time,a noise tolerant multi-label learning model induced by graph trend filtering is proposed,which bridges enriched labels through group sparse constraints,and the Graph Trend Filtering mechanism was introduced to tolerate the inconsistency between the noisy example featuresand labels,thereby reduce the influence of the feature noise on the learning of the enhancement matrix,so as to solve the problem of missing mark in the mixed noise scene.Finally,experiments on seven real multi-label data sets also verify the effectiveness of the model.2.For the incomplete multi-view scene with label noise,an incomplete multi-view and multi-label learning model with latent public representation is proposed,which extends single-view multi-label learning to multi view multi-label learning model.In the case of incomplete view,a latent public representation was learned.At the same time,the model bridges enriched labels through group sparse constraints,so as to solve the problem of missing labels in the scene with label noise.Finally,the experimental results on three multi-view and multi-label data set also verify the predictive performance of the model.3.A large number of formatted sample data manually labeled provide support and guarantee for machine learning algorithms.However,as the cost of labeling increases,new challenges are presented to the task of data annotation.In order to meet the labeling requirements,this thesis designs and implements an image annotation system that uses the multi-label learning algorithm proposed in this thesis to realize image data automatic annotation.