Robust multi-tag feature selection method considering feature label dependency

Abstract

The invention discloses a robust multi-tag feature selection method considering feature label dependency. The method comprises the following steps: firstly constructing a sparse least square regression-based model, and introducing both a manifold regular term and a non-negative matrix factorization term to form a target function; secondly, performing iterative solution operation on the target function to obtain a trained feature weight matrix; carrying out weighting operation of feature tag redundancy on the feature weight matrix, and embedding the feature tag redundancy into the feature weight matrix; and finally, ranking row norm values of the feature weight matrix from high to low, and selecting k optimal features of the optimal features. According to the method, the problems of label redundancy and data noise can be effectively solved, and meanwhile, the information retention degree of the feature weight matrix is improved, so that the model can select feature subsets with higher discriminating power and information power.

Description

Technical field [0001] The present invention relates to machine learning feature selection and pattern recognition, the specific technique is a robust method for selecting multi-label dependence characteristic feature tag consideration. Background technique [0002] Multi-label data corresponding to the plurality of samples is often a label, tag data has more complex than single tag information. In order to solve the data dimension too large caused by the "curse of dimensionality" problem, multi-tag feature selection method came into being. Least squares regression as a common embedded feature selection method, in recent years, has also been used on many labels feature selection method, and often sparse regularization term in conjunction with, the tab forming the sparse feature selection model, which enables feature weighting matrix row sparse, thus guide the selection of a subset of features even more discriminative. In addition, manifold learning has often been used in the leas...

AI Technical Summary

Problems solved by technology

However, it is difficult for existing methods to comprehensively consider the correlation between features and labels, so that some key information is lost. Not only that, noise points and wrong labels are also a problem that needs to be solved.

Images