Structure Preserving Unsupervised Feature Selection Method Based On Autoencoder

With the rapid development of artificial intelligence and big data in various fields,the amount of data in the Internet is also increasing exponentially.These massive amounts of high-dimensional data pose severe challenges to traditional machine learning and statistical analysis theories.The feature space of the original high-dimensional data usually contains many redundant features,background noise,abnormal information,and irrelevant information.If the data is not preprocessed,the use of high-dimensional data for machine learning will seriously increase the time and space complexity,and even encounter"Dimensional disaster"and"overfitting"problems make machine learning tasks impossible.Therefore,feature selection has attracted more and more attention from researchers and engineers,and has become a hot research direction in the field of machine learning.This paper introduces in detail the research status of unsupervised feature selection algorithms based on deep learning framework,analyzes the advantages and disadvantages of different types of unsupervised feature selection algorithms based on deep learning,and aims at these advantages and disadvantages,this paper proposes two new types of unsupervised feature selection algorithms.The method of supervising feature selection is as follows.(1)This paper proposes an unsupervised feature selection algorithm based on autoencoder and manifold regularization.The proposed algorithm selects the optimal feature subset based on the contribution of a single feature to the reconstruction of the original space and the reconstruction of other features.L_2/1 regularization penalizes the weight matrix to make it more sparse,introduces manifold regularization to mine the local and non-local geometric structure of the data space,so that the model learns key information from the original data distribution,and finally constructs a new The objective function,and then use the gradient descent algorithm to optimize the objective function to obtain the optimal parameter solution,and finally obtain the optimal feature subset according to the weight matrix.The experimental results on six public data sets verify that the proposed method can significantly improve the classification accuracy and clustering accuracy compared with other unsupervised feature selection.(2)This paper proposes an unsupervised feature selection algorithm based on autoencoder and adaptive dynamic graph.In order to avoid the pre-defined graph matrix and overcome the model's dependence on fixed similarity,adaptive dynamic graph learning is introduced to unify graph construction and feature learning into an optimization process.Construct structure-preserving graph regularity,introduce manifold regularity to preserve the local and non-local geometric structure of the original data space,and finally construct a new objective function.Using an iterative optimization algorithm,the proposed new objective function is optimized to obtain the optimal parameter solution,and finally the optimal feature subset is obtained according to the weight matrix.The experimental results on six public data sets verify that the proposed method can significantly improve the classification accuracy and clustering accuracy compared with the unsupervised feature selection algorithm based on the graph autoencoder.In summary,this paper proposes two unsupervised feature selection algorithms based on autoencoders for structure preservation.Through the analysis of experimental results,it is verified that the proposed algorithm can effectively remove redundant and irrelevant features and select the most representative feature set.Feature subsets to improve the robustness of subsequent learning tasks.