Fractal autoencoder with redundancy regularization for unsupervised feature selection  

作　　者：Meiting SUN Fangyu LI Honggui HAN 

机构地区：[1]School of Information Science and Technology,Beijing University of Technology,Beijing 100124,China [2]Beijing Key Laboratory of Computational Intelligence and Intelligent System,Beijing 100124,China [3]Engineering Research Center of Digital Community,Ministry of Education,Beijing 100124,China [4]Beijing Institute of Artificial Intelligence,Beijing 100124,China

出　　处：《Science China(Information Sciences)》2025年第2期85-98,共14页中国科学(信息科学)(英文版)

基　　金：supported by National Key Research and Development Project(Grant Nos.2022YFB3305800-5,2023YFB3307300);National Natural Science Foundation of China(Grant Nos.62125301,62373014,92267107);Beijing Youth Scholar(Grant No.037)。

摘　　要：Feature selection is a crucial step in data preprocessing because feature selection reduces the dimensionality of data by eliminating irrelevant and redundant features.Since manual labeling is expensive,unsupervised feature selection has received increasing attention in recent years.However,existing unsupervised feature selection methods tend to prioritize selecting highly correlated features over exploring feature diversity.Thus,a regularized fractal autoencoder(RFAE)method is proposed to select informative features in an unsupervised way.Specifically,the fractal autoencoder network extends autoencoders to construct a correspondence neural network and a selection neural network.The correspondence neural network exploits interfeature correlations and the selection neural network selects the informative features.A redundancy regularization strategy consists of a redundancy elimination regularization term based on the dependency between features and a sparse regularization term based on the group lasso.The redundancy regularization strategy eliminates feature subset redundancy and enhances network generalization ability.Extensive experimental results on six publicly available datasets show that the proposed RFAE outperforms the compared methods regarding clustering accuracy and classification accuracy.Moreover,the proposed RFAE achieves acceptable computation efficiency.

关 键 词：unsupervised feature selection fractal autoencoder correspondence neural network selection neural network re-dundancy regularization strategy 

分 类 号：TP1[自动化与计算机技术—控制理论与控制工程]