Low Rank Feature Selection Algorithms Via Feature Self-representation

With the development of artificial intelligence and big data industries,a lot of high-dimensional data have been accumulated in every domain of human society,and it is urgent to effectively preprocess these high-dimensional data.In order to improve the performance and efficiency of various algorithms in high-dimensional data mining,we must conduct dimensionality reduction preprocessing on high-dimensional data.Therefore,dimensionality reduction has naturally achieved tremendous development and has received extensive attentions.In general,dimensionality reduction methods can be divided into feature selection and subspace learning.Feature selection usually uses some specific models to extract some necessary features from the data,so as to achieve the purpose of reducing the data dimension.Subspace learning projects high-dimensional data into a low-dimensional space by a projection matrix to maintain the correlation structure between data.In short,feature selection is more interpretive than subspace learning,and subspace learning is more stable than feature selection.This paper combines the above two dimensionality reduction methods,aiming at the characteristics of high-dimensional data with more outliers,reasonably using the feature self-representation to relief the impact of outliers,and according to low-rank representation and sparse representation techniques to further remove noises,selecting the most representative feature subset,and finally applying it on the text and face data classification research.The core contents and the original innovative points of this paper are listed as follows:Since feature self-representation has good characteristics while constructing models,this paper combines related techniques such as low rank and hypergraph,put forward an unsupervised feature selection algorithm,low rank hypergraph feature selection algorithm based on self-representation(BHSLR_FS).First,this thesis considers the self-representation matrix to sparsely represent each feature by a linear combination of other features.Such self-representation property is then enforced a low-rank assumption to learn the low-rank representation of high-dimensional data,via considering the global structure of the data to conduct subspace learning.Second,this thesis considers the local structure of the data by a hypergraph based regularizer.In this way,the proposed method integrates subspace learning intothe framework of feature selection.By comparing the experimental results with the comparison algorithms on six real datasets,BHSLR_FS has better classification performance than all the comparison algorithms after conducting feature selection for data.Due to structure learning may provide complementary information to enhance the performance of feature selection,so this thesis proposes a new dimensionality reduction method Unsupervised Feature Selection via Local Structure Learning and Sparse Learning(LSS_FS for short).Specifically,this thesis first utilizes feature self-representation to construct the model.After that,the self-representation coefficient matrix is dynamically adjusted to the optimal state based on the similarity matrix.Then,this thesis uses low-rank representation to explore the global manifold structure of the data.Finally,this thesis combines sparse learning to conduct feature selection.Experimental results demonstrated that the LSS_FS algorithm can select the best discriminative features and achieve the best classification performance,compared to the competing methods.This dissertation mainly focuses on different types of high-dimensional data(including text data and face data),and designs two novel dimensionality reduction algorithm.Specifically,this paper uses the feature self-representation and low rank as the core technologies,and according to different types of data structures,combined with related technologies to select a representative feature subset.At the same time,in order to ensure the fairness of the experiments,all the algorithms in the paper are verified and analyzed in the same experimental environment.In addition,this paper employs three evaluation indicators to verify the effect of the selected feature subset for classification.Experimental results on multiple public datasets show that the new algorithm proposed in this paper is stronger than all the comparison algorithms in terms of robustness and feature selection performance.In the future research,I will consider to integrate deep learning and other related technologies to further optimize and enhance the performance of the proposed dimensionality reduction algorithms.