Unsupervised Feature Selection Methods Based On Data Structure Learning And Particle Swarm Optimization

With the coming of big data era,high-dimensional data become more and more common in real-life.High-dimensional data can save more useful information while they introduce many irrelevant/redundant attributes(features).In order to reduce the time of system learning and improve learning accuracy,how to select representative features has become one of the current research focuses.Compared with supervised and semi-supervised feature selection,unsupervised feature selection becomes very difficult because of the loss of label information.Based on the methods/technologies of data structure learning and particle swarm optimization,this thesis studies efficient unsupervised feature selection algorithm for high-dimensional data,which mainly includes the following three parts:(1)Firstly,by transforming an unsupervised feature selection problem into a supervised one,a feature selection model based on subspace learning guided by non-negative Laplace embedding and its solving algorithm are proposed for maintaining the local structure of data.The proposed method uses the non-negative Laplacian embedded to generate pseudo labels for ensuring classification accuracy;and a feature selection model based on subspace learning is established by fusing the label information generated.After that,an iterative algorithm is given to solve the model,and the convergence of the iterative algorithm is analysed.The proposed method is applied to several typical data sets such as face image process,and experimental results prove its validity.(2)Secondly,to maintain both the global and local structures of data,an unsupervised feature selection model merged with structural learning and its solving algorithm are proposed.In the model,the sparse representation and the probabilistic neighborhood relationship learning are used to learn the global and local structures of data,respectively;next,an unsupervised feature selection model merged with structural learning is presented by using the row sparse characteristic transform matrix to select key features;after that,an iterative solution algorithm that combins structural learning with feature selection is presented.The proposed algorithm is utilized for several typical data sets such as speech data processing,and experimental results demonstrate its superiority.(3)Finally,by combining the global search capability of particle swarm optimization and the local exploitation capability of filter-based approach,a two-stage hybrid unsupervised feature selection method is proposed.In the first stage,the conception of information entropy is used to evaluate the average correlation between features,and a reduction strategy of feature space based on the maximum entropy principle is introduced to remove irrelevant features fast.In the second stage,an improved bare-bone particle swarm optimization is used to selelct key features from the reduced feature space in order to produce high-quality feature subset.Furthermore,a local search operator based on feature redundancy is given to improve the exploitation capability of the swarm.Experimental results on several typical data sets confirm the superiority and effectiveness of the proposed algorithm.