Research On Multi-Label Classification Based On Feature And Label Dual-Space Smoothing

Multi-label classification has been widely used in many fields such as text classification,image classification and so on.However,the existing multi-label classification methods face great challenges in dealing with high-dimensional input features and output label space.Therefore,the focus of this thesis is to solve the dimension disaster caused by high-dimensional data sets,build an accurate multi-label classification model,and improve the accuracy of classification.To solve the above problems,this thesis constructs a multi-label classification model based on feature and label dual-space smoothing.The specific work is as follows:(1)Aiming at the problem that the mainstream multi-label classification algorithms face the disaster of dimensionality in high-dimensional data sets,this thesis proposes a multi-label classification method with geometry preserving dual projections.Considering the highdimensional data of feature and label space,this method seeks to learn from the highdimensional input feature and output label space simultaneously to the dual projections of lowdimensional space.It not only solves the problem of high feature dimension,but also solves the problem of noise and inaccurate classification caused by high label dimension.In addition,considering that the data in the feature and label space are manifolds,we can explore and learn their geometric structure by constructing the Laplace diagram of the label and feature space,so as to further learn the low-dimensional structure.Therefore,a better low-rank structure is realized in label and feature space by constructing dual projections,which further improves the effect of multi-label classification.The experimental results show that the multi-label classification method with geometry preserving dual projections proposed in this study has higher accuracy compared to the comparison algorithm on different types of data sets.(2)On the basis of the above methods,in order to achieve better effect of feature and label projection,this thesis further introduces a manifold structure that maintains the global geometric structure on feature and label projection,that is,an adaptive Laplacian structure is proposed.On the basis of dual projections matrix,we can learn their geometric correlation structure by constructing a common Laplace diagram.Unlike the previous method,this method adaptively constructs a Laplace map directly from the original data.Through continuous updating and learning,it assigns appropriate neighbor data points for each map,better maintaining the geometric structure of the dual projections in the low-dimensional space,and further improving the accuracy of multi-label classification.The experimental results show that the multi-label classification method proposed in this study based on adaptive Laplacian graph structure has higher accuracy compared to the comparison algorithm on different types of data sets.(3)On the basis of the above research,a music recommendation system is designed and implemented.The system uses the current mainstream technology architecture to realize the main functions of the music system.At the same time,the algorithm model studied in this thesis is applied to provide music recommendation functions for users,fully demonstrating the use value of the algorithm proposed in this thesis.