Research On Dual Reconstruction For Feature And Instance Selection

In today's era of data explosion,data dimensionality reduction is one of the common problems in machine learning,pattern recognition and data mining.With the rapid development of data acquisition technology,the emergence of large-scale high-dimensional data is emerging,which generally exists in the process of data processing applications.Today,the development level of data acquisition technology is still limited.There are many kinds of data sources,and the data types are also more diversified.In the acquisition stage,some low-quality data,such as noise data,will be included in the acquisition results.Although high-dimensional data has more comprehensive characteristic information,which can better describe the internal structure of data,it will undoubtedly bring negative effects,such as higher operation and storage costs,dimension disaster and so on.Therefore,in the face of the problem of high-dimensional data processing,many scholars are committed to using feature selection method to achieve data dimensionality reduction.Aiming at the technical difficulties of feature selection to be improved,this paper proposes two corresponding improved algorithms for the effective integration of instance selection and feature selection:(1)An unsupervised dual feature and instance selection algorithm is proposed.This method effectively integrates the feature selection process and instance selection process into a learning framework.From the point of view of instance selection,we use the data points after feature selection to further select instance to ensure that the data points retained by feature selection can well reconstruct the original structure of the data.Similarly,for the feature selection level,after the instance screening process,we select the instance with the largest amount of information to form the instance set.Through the joint action of these two selection tasks,better performance can be achieved at both feature selection level and instance selection level.Through the performance comparison experiments of several benchmark datasets,it is proved that this method is effective and better than many mainstream algorithms.(2)A dual selection method based on local structure preserving strategy is proposed.Based on the unsupervised dual feature and instance selection method,this method considers the importance of local manifold structure to construct the original data structure,and introduces a local regularization operator to extract the local discriminant structure of feature.A more accurate local feature preserving structure is used to further improve the quality of feature selection results.At the same time,in the integrated bi-directional selection framework,the performance improvement of feature selection will also bring positive effects for the instance selection process.Through the performance comparison experiments of several benchmark datasets,it is proved that this method is effective and better than many mainstream algorithms.An unsupervised feature selection algorithm analysis system is designed and developed.The system includes three modules: initialization module,unsupervised dual instance and feature selection algorithm module and algorithm result display module.The system integrates the experimental results of the unsupervised dual and feature and instance selection algorithm with its 13 comparison methods.Users can choose the performance evaluation index and display form to display the performance.