Feature Selection Method Based on L21 Normal Form Distance Metric

Abstract

The invention relates to the field of data preprocessing, and discloses a feature selection method based on L21 normal form distance measurement, comprising the following steps: inputting the original data matrix and parameters; performing SVD decomposition on the original data matrix to obtain a column-orthogonal matrix; using the original data matrix Calculate the inter-class scatter matrix and the intra-class scatter matrix in the projection space with the column orthogonal matrix; calculate the defined diagonal matrix D; add the diagonal matrix D and the intra-class scatter matrix, and then compare the inter-class scatter degree matrix to obtain a new matrix; perform eigenvalue decomposition on the new matrix to obtain the eigenvector corresponding to the minimum eigenvalue; update W; iterate steps 3 to 7 until W converges; calculate each row vector of W after the convergence is completed The two-norm value of , select the characteristic number. The invention can greatly improve the robustness and sparseness of the feature selection algorithm, so that the feature selection algorithm can eliminate redundant features and noise features and select the most representative features.

Description

technical field [0001] The invention relates to the field of data preprocessing, in particular to a feature selection method based on L21 paradigm distance measure. Background technique [0002] As one of the data preprocessing methods, feature selection, which can reduce the data dimension, retain the semantic features of the original data, and express high-dimensional data with low-dimensional data, has attracted more and more attention from academia and industry, and has been used in human face Specific application fields such as recognition, image processing, and big data cleaning play an increasingly important role. In recent years, M.Masaeli et al. proposed a linear discriminative feature selection method based on infinite-paradigm distance (Masaeli, M., G. Fung, and J.G. Dy. From Transformation-Based Dimensionality Reduction to Feature Selection. in International Conference on Machine Learning. 2010.) has attracted much attention for its robust noise immunity. The p...

AI Technical Summary

Problems solved by technology

However, since the algorithm will obtain a trivial solution under certain circumstances, the algorithm loses the ability of feature selection. In 2016, Hong Tao et al. proposed a discriminative feature selection method through the L21 norm (Hong, T., et al., Effective Discriminative Feature Selection With Nontrivial Solution. IEEETransactions on Neural Networks & Learning Systems, 2016. 27(4): p. 796-808.), it is recommended to replace the infinite 1 norm with the L21 norm, which can avoid the existence of trivial solutions

[0003] However, the above feature selection algorithms are only improved in the regular term, which can only provide a certain degree of sparsity and robustness

Images