Advanced Research Of Feature Selection And Spectral Clustering Based On Graph Learning Theory

The graph method is widely used in different fields because it can calculate and maintain the inherent relationship of data compared with the original data and can show a stronger expression.Especially in the field of machine learning,the structure-preserving nature of the graph can ensure that the original structure of the data remains unchanged when more effective information is obtained during the learning process.Among the different graph structure preservation methods,the graph's local neighbor relationship preservation method has been applied in spectral feature selection algorithm and spectral clustering algorithm.However,the local structure construction method of the previous graph local neighborhood relationship maintenance method only relies on the Euclidean distance to measure the sample similarity relationship in space.Once the noise or redundancy in the data will affect the quality of the established graph matrix and further affect the final Machine learning model learning effect.Therefore,in this paper,we will use two different improvement strategies to propose two methods that can establish a higher quality graph matrix and use the new method to propose more effective graph learning-based machine learning algorithms for the problems of existing graph learning.The main part of the paper is as follows:?1?Spectral clustering algorithm?LCSC algorithm?based on local covariance and regularization.The LCSC algorithm will combine graph learning,local covariance and data regularization to propose an efficient spectral clustering learning model.This algorithm solves the problem of clusters junction caused by a single Euclidean distance measurement through the local covariance matrix of samples,and uses the regularization method to normalize the similarity level of the samples to improve the accuracy of the clustering algorithm.Specifically,LCSC first adds the distance between sample covariance matrices as a supplementary judgment condition in the traditional clustering algorithm to improve the quality of the similarity matrix,and then uses the regularization method to balance the magnitude of the matrix elements obtained,and finally obtains a more accurate spectrum Clustering algorithm model.Through clustering experiment evaluation,the LCSC algorithm achieves better results on real data sets than other clustering algorithms.?2?Dynamic spectral feature selection algorithm based on spectral rotation strategy?DFS-SR algorithm?.The DFS-SR algorithm will combine a spectral rotation method,graph learning and sparse learning techniques to propose a robust spectral feature selection algorithm model.This algorithm combines self-representation graph learning,new sparse norms and spectral rotation techniques to improve the performance of the feature selection algorithm in processing real data.Specifically,the algorithm first uses the self-representation-based graph learning method to replace the traditional Euclidean distance-based graph learning method to obtain a high-quality spectral matrix;then adds the spectral rotation technology to further improve the quality of model learning by fine-tuning the projection direction of the original data matrix through the real data label;finally adds a sparse regularization norm(l_2,1-norm)with group sparse to the model to improve the effect of the final feature selection.It has been verified by clustering experiments that the algorithm can achieve better results than the comparison algorithm.This thesis firstly reforms and improves the problem that the spectral feature selection algorithm and the spectral clustering algorithm in the traditional machine learning algorithm rely on a single Euclidean distance-based data local structure preservation method,which may cause the poor quality of the graph matrix obtained during the learning process.Then,the proposed method and compared methods are evaluated through real data experiments,in which all feature selection algorithms will use the classical clustering algorithm K-means algorithm as the evaluation method of feature selection effects.In order to verify the correctness and effectiveness of the proposed algorithm,the paper uses multiple evaluation indicators to verify and analyze the proposed algorithm results,and all the algorithms in the paper will be tested under unified experimental conditions.The final experimental results show that the new algorithms proposed in this paper are superior to the selected similar algorithms.In future work,I will consider applying the proposed graph learning improvement method directly to classification,regression,or more real application scenarios.