Label-specific Feature Multi-label Learning Based On The Combination Of Multiple Correlation Information

Multi-label learning is one of the main frameworks for processing objects with rich semantics in the real world,which has been successfully applied in text classification,information retrieval,image processing,and other fields.In traditional single-label learning,each sample is labeled by only one label,and the labels are independent of each other.In multi-label learning,each sample is labeled by several labels and there is a certain correlation among the labels.The goal of multi-label learning is to learn from the known multi-label data set and build a model to predict the label set of unknown samples.With the increase in the number of labels,the expression of multi-label data sets becomes more and more diverse,and multi-label learning faces a series of new challenges.Multi-label learning mainly faces three challenges: Firstly,different class of labels in real life should have their own unique features,which enrich the hidden information of labels.Mining the specific features of labels can greatly improve the classification effectiveness.Secondly,the number of labels in multi-label learning is often large,which leads to an exponential increase in the prediction difficulty of unknown samples.In real life,there is a certain correlation among labels.Therefore,mining the correlation information among labels is helpful to reduce the complexity of the algorithm and improve the classification effectiveness.Finally,since most data sets are non-uniformly distributed but tend to be skewed in real life,the number of samples of a certain class is much larger than that of other classes,which ultimately leads to the bias of the classifier.In other words,multi-label learning faces a serious label imbalance problem.This thesis focuses on these three challenges and proposes a label-specific feature multi-label algorithm LSF-MI(Label-specific Feature Multi-label Learning Based on the Combination of Multiple Correlation Information)based on the combination of multiple correlations.The main research contents are as follows:On the issue of how to use the association information among labels to construct labelspecific features,an algorithm that combining label correlation,instance correlation and feature correlation is proposed.The label correlation is calculated by cosine similarity,the instance correlation and feature correlation are calculated by k-nearest neighbor probability graph model,and finally the running time is saved by accelerating the proximal gradient method.The comparative experiments of six multi-label algorithms show that the algorithm performs well in the evaluation metrics on the whole,which proves the effectiveness of the LSF-MI algorithm.In order to verify that the LSF-MI algorithm can be used as a feature selection method of multi-label learning or a general strategy to improve multi-label classification algorithm,a label denoising model is added to the LSF-MI algorithm.Firstly,the original label association set is assumed to be a candidate label set.Because it is affected by noise,the multiple class labels associated with each training sample are only partially valid candidate labels.By combining the candidate label set with the similarity graph matrix,the noise is eliminated and the real label set of each training sample is estimated.The label-specific features of the denoised LSF-MI algorithm are extracted and applied to three classical binary classifier algorithms,and relevant experiments are done on seven data sets.The experimental results show the effectiveness of the label denoising model algorithm and the feasibility of the LSF-MI algorithm as a general feature selection method.In addition,the LSF-MI algorithm as a feature selection method can alleviate the problem of label imbalance to a certain extent.