Research Of Dimensionality Reduction And Clustering Based On Constraint Weight Learning And Dictionary Learning

With the rapid development of information technology,all walks of life have produced a large number of high dimensional data,how to mine from the high dimensional data is a focus of modern machine learning research.However,the high dimensional data storage is huge and the computational cost is very high,which make the traditional machine learning algorithm face great challenge.This is what we call ”curse of dimensionality”.As an effective way to solve the problem of learning from high dimensional data,dimensionality reduction and subspace clustering have been widely concerned by researchers.Both dimensionality reduction and subspace clustering are the process of simplifying the expression of data;dimensionality reduction aims to preserve the discriminative characteristics,and exclude those common information,it can be seen as a special clustering method.Similarly,subspace clustering aims to group data into a few clusters,which also can be seen as a special dimensionality reduction method.With the accumulation of high dimensional data,both of them have been extensively studied and widely applied.Following discussion on the development history and existing challenges of semisupervised dimensionality reduction method and subspace clustering method,this dissertation focus on these topics,i.e.weighted pairwise constraints,graph construction and optimization,dictionary learning,latent space learning,and its applications in dimensionality reduction and subspace clustering.The main results and contributions of this dissertation are as follows.1.An adaptive semi-supervised dimensionality reduction method based on pairwise constraints(ASSDR-PPC)is proposed.This method utilizes supervised information by weighting the constraint with probability,and preserves the inner structure of data by constructing sparse graph.It also merges projection matrix computation,graph construction and constraint weight optimization into a whole to guide dimensionality reduction.Extensive experiments show that our proposed method performs better than many other semi-supervised dimensionality reduction methods.2.Low-Rank and Sparse Subspace Clustering based on Latent Space Dictionary Learning(LRSSC-LSDL)is proposed.We introduce dictionary learning to improve the ability to express data,and learn projection to accelerate the speed of computing sparse and low-rank representation.Extensive experiments show that our proposed method performs better than many existing state-of-the-art self-expressiveness based subspace clustering methods when there is a heavily data corruption.3.Semi-Supervised LRSSC-LSDL(SSLRSSC-LSDL)is proposed,which is a semisupervised version of LRSSC-LSDL.This method utilizes supervised information by weighting the constraint with a probability.It also can simultaneously learn dictionary,projection matrix and weight constraint to get low-rank and sparse representation for subspace clustering.Extensive experiments show that our proposed method performs better than many existing subspace clustering methods and semi-supervised clustering methods.