Research On Multi-label Classification With Incomplete Label Information

Multi-label learning is ubiquitous in practical applications and it is one of the main problems in the field of machine learning.Unlike traditional single-label learning,multi-label learning can handle instances with multiple class labels,at the same time,it can predict a set of corresponding labels for a new instance to be tested at the testing stage.For multi-label learning,a large number of training data with complete label information is a prerequisite for better learning performance.However,in practical problems,because of the large scale,various types and difficult labeling of data,the training data obtained are often a large number of weakly labeled data with incomplete label information.How to use these weak labeled data to perform classification tasks and achieve better classification results has gradually become one of the hot issues of research.In this paper,two different types of weakly labeled data encountered in the process of multi-label classification are studied.The main works are as follows:Firstly,for the weak labeled data in which only part of relevant label information are labeled in training data,this paper proposes a multi-label classification algorithm suitable for this type of weakly labeled data.The algorithm recovers incomplete label information by iteratively updating the weight of each instance and utilizing the correlation between any two labels.Then after the labels are recovered,a classification model is trained using the training set with complete label information,and the model is used to predict testing set.In the experimental stage,the algorithm is simulated with other related algorithms on several public multi-label datasets.Experimental results show that the proposed algorithm achieves better results in both label recovery and final multi-label classification.Secondly,for the weak labeled data with partially labeled and completely unlabeled instances in training data,this paper proposes a semi-supervised multi-label classification algorithm suitable for this type of weakly labeled data.The algorithm first recovers the missing label information in partially labeled instances completely.Then the label information of the recovered instances is transmitted to the completely unlabeled instances by the way of label propagation.After that,a graph-based semi-supervised linear classifier is trained with all instances(including training instances and testing instances),and the classifier is used to predict the corresponding label information of testing instance.In the experimental stage,the algorithm is simulated on several public multi-label datasets,and compared with the existing related algorithms,then the effectiveness of the proposed algorithm is illustrated.Finally,in this paper,the research contents and achievements are briefly summarized,and the next stage of research is planned and prospected.