Research On Semi-supervised Multi-label Propagation Algorithm Based On LTSA

With the development of the information and technology,data has changed from the traditional simple structure and small-scale data into a large scale information carrier with complex structure and high dimension.Compared with traditional single label learning,multilabel learning can process richer information.Therefore,multi-label learning has been widely used in machine learning,data analysis,pattern recognition and other fields.However,in the real world,the data with labels are very rare,so semi-supervised multi-label learning is a very important research topic.In addition,there is often a complex relationship between features and labels in multi-label learning,so how to explore the relationship between features and labels has also become an important research topic.In order to explore the complex relationship between labels and features,and complete label propagation under semi-supervised situation,this paper combined with the local tangent space alignment algorithm(LTSA),proposed the local tangent space alignment based on semisupervised multi-label information(SSLTSA).And then proposed the semi-supervised multilabel propagation based on LTSA(SSLTSAMLP).At the same time of dimensionality reduction of data features,linear mapping is established between low-dimensional feature space and label space.Combined with hypergraph Laplacian matrix,semi-supervised label propagation is completed.Considering that manifold learning can only depict the local manifold structure of data and cannot depict the global structure of data,this paper integrates the subspace low-rank representation on the basis of the semi-supervised multi-label propagation based on LTSA and subspace representation,and proposes a semi-supervised multi-label propagation based on LTSA and low-rank subspace representation(SSLTSALRRMLP).The semi-supervised label propagation is completed under the manifold structure of guaranteed data.In this paper,several different evaluation indexes are selected,and the two proposed methods are compared with other representative methods to verify the superiority of the proposed method in semi-supervised multi-label learning.