Research On Multi-label Classification Algorithm Based On Class-specific Features And Dependent Labels

At present,there are a large number of objects that contain a variety of semantic information in the Internet,and the traditional single-label learning framework has been difficult to effectively deal with these multi semantic objects,while the multi-label learning framework can deal with these semantic objects well,which has attracted the attention of many researchers.In multi-label learning,the problem of classification is a hot topic discussed by current researchers.Different from single-label classification,the output space of the multi-label classification problem corresponds to multiple labels,and there may also be some implicit correlations between the labels.If we can make full use of the correlation between labels,it will be of great help to improve the classification performance of the algorithm.For example,some multi-label classification methods utilize high-order correlations between labels by constructing chain classifiers or stacking classifiers.but these methods may have the problem of label prediction error propagation in the testing phase.In addition,in the multi label classification task,features are an important reference for describing different data objects in the data set.However,redundant or irrelevant features will not only increase the computational cost,but also affect the classification performance of the algorithm.Focusing on the problems to be solved in multi-label classification,the main research work of this paper is as follows:A multi-label classification algorithm based on class-specific features and dependent labels is proposed.The algorithm constructs sparse class-specific features coefficient matrix and dependent labels coefficient matrix through7)₁ sparse regular terms.The construction of class-specific features can select the most discriminative feature subset of each category label,and the dependent labels can reflect the global correlation between labels.At the same time,combined with the concept of local correlation of labels,the coefficient matrix of the model is constrained by manifold regular terms to limit the label output.The algorithm solves the objective function by accelerating gradient descent method.In the test phase,the predictive labels can be obtained by solving the class-specific features coefficient matrix and the dependent labels coefficient matrix without using additional classifiers,so as to reduce the impact of error propagation caused by using external classifiers on the classification performance of the algorithm.In the experimental stage,the classification performance of the proposed algorithm is compared with other multi-label classification algorithms using different scale datasets,and the superiority of the classification performance of the proposed algorithm is verified according to the evaluation results on different evaluation metrics and the subsequent algorithm performance test results.