Nonnegative matrix factorization method based on discriminative orthogonal subspace constraint

Abstract

The invention discloses a nonnegative matrix factorization method based on a discriminative orthogonal subspace constraint. The method mainly comprises the following steps of (1) stretching an image in a training sample set into vectors to compose a training data matrix Xtrain, then factorizing the Xtrain in a nonnegative matrix factorization framework based on the discriminative orthogonal subspace constraint, and directly exerting a discriminative constraint item based on within-class and between-class associations to the basis matrix; (2) constructing a projection matrix W by use of the learned basis matrix U*, calculating projection expression of the training data Xtrain and test data Xtest in the projection matrix W, and performing an image recognition experiment with a nearest neighbor classifier; and (3) calculating the image identification precision. According to the nonnegative matrix factorization method based on the discriminative orthogonal subspace constraint, the discriminative structure information inside the data are explored and utilized, the discriminative constraint directly exerted to the basis matrix in the algorithm enhances the generalization performance of the algorithm and improves the image identification effect; and the method can be widely applied to the field of data mining and data analysis.

Description

technical field [0001] The invention belongs to the technical field of information processing, and in particular relates to a non-negative low-dimensional data processing method, which can be used for data mining, data analysis and the like. Background technique [0002] Non-negative matrix factorization, as a feature extraction technique, is widely used in clustering and classification tasks. For unsupervised clustering tasks, data distribution information can usually be used to improve the performance of non-negative matrix factorization algorithms, so that the extracted features have better representation capabilities. For supervised classification tasks, the data category label information can be used to encode the discriminative structure information of the data, which can assist the non-negative matrix factorization algorithm to learn features with better classification ability. [0003] According to the different application methods of the labels, the supervised non-...

AI Technical Summary

Problems solved by technology

[0006] Most supervised non-negative matrix factorization algorithms impose discriminative constraints on the encoding matrix, however, when processing test samples, the base matrix is ​​usually used to construct the projection matrix, but since the connection between the discriminative information and the base matrix is ​​indirect, Therefore, the generalization ability of the algorithm on the test data set is not good enough

Images