Study On Multi-label Specific Feature Selection Based On Mutual-information

Feature selection is an extremely important step in pattern recognition and data mining.It can transform the original sample features from the high-dimensional to the low-dimensional feature space,thus greatly alleviating the problem of the dimensional disaster in the field.So far,the feature selection methods in the field are mainly focused on single-label learning,but few methods of feature selection for multi-label learning;Even if there exists a small number of multi-label learning algorithms,it is possible to achieve the purpose of learning based on the same feature space in different class labels.Actually,different class labels may have their own specific features in multi-label learning algorithms.Therefore,this paper will mainly start from the following aspects to study the multi-label feature selection algorithm.First of all,this paper discusses the analysis of feature selection algorithm in the current field,summarizes the classification,study the literature carefully,considers shortcomings existing in the current algorithm,and proposes the corresponding improvement measures.For instance,most feature selection algorithms in this field is based on single-label learning,while in real-life,a lot of samples have multiple class label,and mostly exist in the form of overlapping between the class label.Therefore,this paper will study the multi-label feature selection algorithm.Second,the author noted that the feature selection method proposed in the current multi-label learning in the field is based on the same sample feature space under different class labels to achieve the purpose of learning,which may contradict the fact that in multi-label learning algorithm,different class labels may have their own specific characteristics.Therefore,this paper also studies label-specific features of multi-label learning.In addition,in the study of the label-specific features of multi-label learning,employing the mutual information to measure significance of sample features and effectively select out the significant features for classification discriminant from the label-specific features of multi-label learning,and remove the redundant attributes.In the end,the author also extensively reviewed the main location of multi-label samples.The study found that the multi-label samples were mainly on the boundary.Hence,in order to improve operation efficiency of the algorithm,this article employed the sample selection techniques to select out the boundary samples,thus greatly reducing the computational complexity of the algorithm and the operation time,and maintaining a satisfactory classification performance in the meantime.