URMFFS (unsupervised regularization matrix factorization feature selection) method

Abstract

The invention provides a URMFFS (unsupervised regularization matrix factorization feature selection) method. According to the method, inner product regularization constraint is introduced, correlation between selected features is considered, and accordingly, a selected feature subset can represent original high-dimensional features well and has lower redundancy. The invention further designs an iterative optimization algorithm for solving an objective function of URMFFS. A large quantity of comparison experiments of the URMFFS method and current popular unsupervised feature selection methods are carried out by the aid of six common databases including AR10P, Yale, ORL, Jaffe, PIE10P and TOX-171, and experimental results indicate that the performance of the URMFFS method is remarkably superior to that of other unsupervised feature selection methods.

Description

technical field [0001] The invention relates to the technical fields of signal processing and data analysis, in particular to an unsupervised regularization matrix decomposition feature selection method. Background technique [0002] With the increasing popularity of computer technology, social network informatization, and the Internet, a large amount of high-dimensional data, such as text, multimedia, video, and images, has emerged, and they come from different systems, sensors, and mobile devices. These high-dimensional data usually have the characteristics of diversity, complexity, and redundancy. If computers are used to directly process these high-dimensional data, it will require huge memory storage space and computing costs, and will cause "curse of dimensionality" in severe cases. problem, which reduces the computational efficiency and performance of the algorithm. The "curse of dimensionality" is to ensure a certain accuracy rate in calculations involving vectors, ...

AI Technical Summary

Problems solved by technology

[0013] The matrix factorization feature selection (Matrix Factorization Feature Selection, MFFS) algorithm successfully builds a bridge between matrix factorization and feature selection, and its performance is better than a large number of current feature selection methods, but the constraints of the MFFS method are too strict, and It is difficult to satisfy in practical applications, and also ignores the correlation between features, resulting in a certain redundancy in the selected feature subset rather than the optimal feature subset

Images