Research On Multi-label Learning Method Based On Label Correlation Embedding

In multi-label learning,any object is usually described by an instance and contains multiple labels,and the label may be related to each other.The main purpose of multi-label learning is to establish a classifier based on the training data of existing label to predict related label sets for new instances.As is known to all,using the correlation between labels to build a multi-label classification model can definitely enhance the performance of model.Existing methods mainly model label correlation in an indirect way,i.e.additional constraints are added to the coefficients or outputs of the classification model based on the pre-learned label correlation matrix,and the inherent correlation between different labels is not well maintained.Then,the high dimension of feature space will inevitably produce redundant features,which will increase the computational pressure and memory cost,and bring great challenges to multi-label learning.By exploring label correlations and aiming at the problem of large dimension of feature space,in this thesis,a multi-label learning method by exploiting label correlation embedding is introduced,namely MLLCE,which integrates feature space reduction and label correlation into a unified framework.The method firstly implements feature extraction based on the mapping matrix,and obtains the feature space with reduced dimension.Meanwhile,an embedding matrix is learned from the pre-learned label correlation matrix via way of graph embedding,taking the embedding matrix as the coefficient matrix of the model.Finally,we can learn a mapping relation that can maps the dimension-reduced feature space to the original label space,namely classifier.The classifier can predict relevant class labels for new data.When the data cannot be separated linearly,the classification performance of the model will be reduced.Therefore,this thesis also put forward a nonlinear multi-label learning method based on label correlation embedding,namely NL-MLLCE.First,the original feature space is processed by nonlinear feature map.Meanwhile,the coefficient matrix of the model is obtained by the way of graph embedding.Finally,a multi-label classifier is established,and the multi-label data is effectively classified.Based on fifteen datasets and seven commonly used evaluation metrics,the thesis compares the proposed ways MLLCE and NL-MLLCE with various multi-label learning ways.Experimental results show that the proposed ways has competitive performance classification in multi-label learning.