Unsupervised feature selection method and device based on dictionary and sample similarity graph

Abstract

The invention relates to an unsupervised feature selection method and device based on a dictionary and a sample similarity graph. The invention discloses an unsupervised feature selection method and system based on dictionary and sample similarity graph learning. According to the method and the system, a new dictionary base space of an original data matrix X is provided, and in the generated new dictionary base space W, l2, 1 norm is used for applying row sparsity to W so as to measure the importance of features. Compared with a conventional low-level representation method in an original feature space, the dictionary and sample similarity graph-based learning model disclosed by the invention captures higher-level and more abstract representations, and has a wide application prospect.

Description

technical field [0001] The invention relates to the fields of signal processing and data analysis, in particular to an unsupervised feature selection method and device based on dictionary and sample similarity graph learning. Background technique [0002] With the rapid development of sensors and internet media, high-dimensional analysis and big data have become a challenging and inevitable problem. A large amount of high-dimensional data is used in many applications, such as computer vision, machine learning, pattern recognition, and medical analysis. Although data is usually represented as a high-dimensional feature vector, only a small but unknown subset of features is very important and discriminative for learning tasks. Directly processing these high-dimensional data will not only significantly increase the demand for computational resources, but also degrade the performance of many existing algorithms due to the curse of dimensionality. Feature selection is generally...

AI Technical Summary

Problems solved by technology

First, most existing algorithms perform feature selection on the original feature space. Due to the noise and redundant features of the original feature space, the results are often inaccurate, and higher-level and more abstract representations cannot be utilized.

Second, traditional UFS methods usually construct a similarity matrix and perform feature selection separately. Therefore, in the subsequent process, the similarity matrix obtained from the original data remains unchanged, but the noise samples and features contained in the real data inevitably make the similarity matrix unreliable

This unreliable similarity matrix may destroy the local popularity structure of the original data, making the result less optimal

Images