Research And Application Of Learning Feature Representation Based On Matrix Factorization And Adaptive Graph

In the information age,where data collection and Internet technology are developing rapidly,it is effortless for people to capture massive and high-dimensional data in real life.However,high-dimensional data often contains many redundant features and noisy information,which poses a significant challenge to current data mining and analysis.This requires us to dimensionally reduce the high-dimensional data to learn the most useful information;therefore,feature representation learning becomes crucial.In this paper,we conduct an in-depth study of matrix decomposition and adaptive graphs as follows:(1)This paper proposes a novel unsupervised feature extraction method that aims to improve the quality of feature extraction by reducing the impact of redundant features and noisy information in high-dimensional data on graph learning.The method is named the robust non-negative matrix decomposition for multi-constrained adaptive graph learning(RRNMF-MAGL)method,which can extract robust low-dimensional features and adaptively obtain the flow structure to reflect the data distribution well.Specifically,robust non-negative matrix decomposition(RNMF),multi-constraint adaptive graph learning(MAGL)based on low-dimensional features,and graph Laplace regularization term(GLR)are combined into a unified framework to improve feature representation.In addition,sparse and local constraints are introduced into the MAGL model to enhance the discriminative power of the learned graph structure.In this paper,the effectiveness of RRNMF-MAGL is demonstrated by conducting relevant experiments on eight datasets.(2)This paper proposes a novel unsupervised feature selection method,namely,robust adaptive structure learning based on matrix decomposition(RMFRASL),which aims to select discriminative features from massive data and remove redundant or irrelevant features.RMFRASL integrates robust matrix decomposition-based feature selection(RMFFS),adaptive structure learning(RASL),and structure regularization(SR)into a unified framework to improve the effectiveness of feature selection.By comparing with existing unsupervised feature selection methods,experimental results show that the RMFRASL method proposed in this paper is highly robust and adaptive and can effectively select useful features to improve the accuracy and performance of the model in classification and clustering tasks.(3)This paper proposes a semi-supervised discriminative non-negative matrix decomposition(SDNMF)method to improve the performance of NMF feature representation methods by making full use of the label information of the data.SDNMF integrates soft label NMF(SLNMF),label propagation(LP),and adaptive graph learning(AGL)into a unified framework for effective feature learning and unknown sample label prediction.In this paper,we conduct extensive experiments on standard datasets,and SDNMF can obtain higher classification and recognition performance metrics than related methods.