Local scale parameter, entropy and cosine similarity-based spectral feature selection method

Abstract

The invention discloses a local scale parameter, entropy and cosine similarity-based spectral feature selection method. A Gaussian kernel function is adopted as a similarity measurement method, defined local scale parameters which are of features and based on local standard deviation of the features are used as kernel function parameters, and problems that uniform scale parameters in calculating feature affinity matrices cannot reflect data distribution information, and local scale parameters are impacted by off-group points are overcome; entropy and cosine similarity sorting is respectively adopted to measure feature importance degrees, a suitable feature subset can be quickly selected; and technical support is provided for data analysis of diseases such as tumours, and the method has important biomedical significance.

Description

technical field [0001] The invention belongs to gene microarray data and text data analysis technology, and relates to a spectral feature selection method based on local scale parameters, entropy and cosine similarity. Background technique [0002] Feature selection is the primary task of high-dimensional big data analysis such as gene microarray data and text data [1,2] , whose goal is to remove irrelevant or redundant features from all features, and select a feature subset with good distinguishing ability, so as to preserve all the classification information of the original feature set as much as possible. According to whether the feature selection process uses sample class label information, feature selection algorithms are divided into supervised methods and unsupervised methods [3] . The supervised feature selection method performs feature selection by calculating the correlation between features and class label columns, while the unsupervised feature selection method...

AI Technical Summary

Problems solved by technology

[0036] Aiming at the deficiencies of the prior art, the present invention proposes a spectral feature selection method based on local scale parameters, entropy and cosine similarity, which overcomes the defect that the unified scale parameter σ cannot completely and accurately reflect the data distribution information when calculating the feature affinity matrix , while overcoming the feature local scale parameter σ when the self-tuning method calculates the feature affinity matrix i , the problem affected by outliers

Images