Research On Robust Multi-label Classification Based On Labeland Nearest Sample Space

Multi-label classification is a classification model that assigns a set of related labels to unknown samples,which has been widely used in image retrieval and recommendation tasks.However,the high dimensional correlation relations of samples and labels and noise problems become an important challenge for multi-label classification.For these two issues,this thesis constructed robust multi-label classification method based on label and near-neighbor sample space to improve the accuracy of classification model prediction results.The main research contents of this thesis are as follows:(1)For the traditional multi-label classification model often only consider the feature selection problem in the sample feature space,a robust multi-label classification model based on the label and feature space co-learning is proposed.The main idea is: considering that the traditional multi-label classification method is a regression idea to directly build the relationship between data samples and their labels,which is also viewed as a feature learning method.The thesis further applying it to high-dimensional label space to obtain both low-rank feature space and low-rank label projection simultaneously.Furthermore,since the high dimensional feature and label space are usually affected by different noise injections,the nuclear norm method is introduced to learn the original noise space.In this way,more discriminative features can be effectively obtained from input noisy feature space and output noisy label space,which can further learn the robust low-rank feature and label space.The experimental results show that the multi-label classification method for label and feature space co-learning proposed in this study has higher accuracy compared with the state-of-art multi-label classification models on different types of classification datasets.(2)For the existing multi-label classification model is susceptible to sample noise,this thesis proposed a robust multi-label classification method based on the nearest neighbor sample space.The main idea is: Considering that the samples and their neighbor data have similar semantic structure,the training data of the model is further extended by constructing multiple enhanced datasets consisting of sample neighbors.Specifically,using this data expansion strategy,a more robust subspace is learned not only in the original label and feature space,but also in their multiple expanded nearest neighbor space.Therefore,the proposed method can better suppress noise data from the model,and then learn more robust low-rank features and label projections.Experimental results show that the proposed robust multi-label classification algorithm based on the nearest-neighbor sample space outperforms other state-of-the-art multi-label classification methods.(3)Based on the above robust multi-label classification algorithm,multi-label image classification system is designed and implemented.The operation and test of the system show that the interface is friendly and the function of multi-label image classification.