Abstract:

Due to the so-called curse of dimensionality, which is inevitable and tricky in high-dimensional data analytics, it is of great importance to perform dimensionality reduction via feature selection methods. Therefore, an unsupervised feature selection model based on robust matrix factorization and adaptive graph (MFAGFS) is proposed, which can perform robust matrix factorization, feature selection and local structure learning under a unified learning framework. The model first obtains cluster tags by robust matrix decomposition, cluster tags and local structure information are used to guide the feature selection process. Then, learning the local structure of the data adaptively from the result of feature selection. MFAGFS can accurately capture the structure information of the data and select discriminative features through the interaction between the two basic tasks of the local structure learning and feature selection. Then, the optimization method of the algorithm is described in detail, and the convergence of the algorithm is proved. Finally, experimental comparative analysis and parameter sensitivity analysis are carried out on six public data sets to verify the effectiveness of the proposed model. The experimental result shows that the performance of the proposed methods presented is improved in different degrees compared with other methods.

Key words: feature selection, graph embedding, adaptation, matrix factorization