Linear discriminant analysis dimension reduction method based on cosine similarity weighting

Abstract

The invention discloses a linear discriminant analysis dimension reduction method based on cosine similarity weighting. The method comprises steps: 1, a to-be-acquired initial feature F of each sample in a data set X is read; 2, based on an LLE algorithm, the initial feature F is subjected to initial dimension reduction to acquire a temporary feature F'; 3, sample feature data are acquired, and the temporary feature F' serves as an input feature; 4, a mean value m of each sample class and a mean value m of the total samples in the data set X are calculated; 5, based on the sample feature data, the m and the m, a within-class scatter matrix based on cosine similarity weighting and a corresponding between-class scatter matrix are acquired; 6, an objective function based on the cosine similarity weighting is built to carry out further dimension reduction on the sample feature data; and 7, according to a projection matrix generated in the step 6, the input feature is mapped to new dimension space. The method has better within-class coupling and between-class scatter, and better dimension reduction effects are achieved.

Description

technical field [0001] The invention belongs to the field of data analysis, and in particular relates to a linear discriminant analysis dimension reduction method based on cosine similarity weighting. Background technique [0002] Discriminant analysis is one of the important methods in the field of data analysis, widely used in data classification, target recognition, exception detection, clustering, image processing, biological information processing and other fields; it is mainly based on statistical analysis, according to the distribution of training data itself characteristics, constructing straight lines or curves for segmenting data; the current main methods include linear discriminant method, distance discriminant method, Bayesian discriminant method and Fisher discriminant method, etc. [0003] However, the above classification methods all assume that the samples of each class are distributed in the same aggregation area. For the case where the samples of the same c...

AI Technical Summary

Problems solved by technology

However, it should be noted that the above methods only consider the covariance information between sample vectors and ignore the similarity information of other categories, which cannot fully characterize the degree of dispersion between samples.

Images