随机森林与自编码器相结合的自适应特征选择算法  

Adaptive Feature Selection Algorithm Combining Random Forest and Autoencoder

作　　者：闫蒙蒙 陈建凯 孟会贤[1] 王鑫 YAN Mengmeng;CHEN Jiankai;MENG Huixian;WANG Xin(School of Mathematical and Physical,North China Electric Power University,BeiJing 102206,China;School of Control and Computer Engineering,North China Electric Power University,Beijing 102206,China;Hebei Key Laboratory of Machine Learning and Computational Intelligence,College of Mathematics and Information Science,Hebei University,Baoding Hebei 071002,China)

机构地区：[1]华北电力大学数理学院,北京102206 [2]华北电力大学控制与计算机工程学院,北京102206 [3]河北大学数学与信息科学学院河北省机器学习与计算机智能重点实验室,河北保定071002

出　　处：《人工智能科学与工程》2023年第9期39-47,共9页Journal of Southwest China Normal University(Natural Science Edition)

基　　金：2022年度河北省社会科学发展研究课题(20220303038)。

摘　　要：特征选择在信用评估中是一种常用的数据降维技术。然而,传统的特征选择方法主要基于特征之间的线性相关性,无法有效处理非线性数据关系,导致无法准确估计变量之间的相关性程度。为了克服这个问题,该文提出了一种改进的特征选择算法,结合了随机森林和自编码器的优点。首先,利用随机森林去除与目标变量不相关的特征。然后,计算剩余特征的综合重要度,并使用这些保留的特征来训练自编码器神经网络。最后,使用自编码器的学习参数初始化一个三层神经网络,用于重构特征。在公开的信用评估数据集上进行了实验,结果表明,所提出的算法相对于其他方法表现更出色。Feature selection is a commonly used data dimensionality reduction technique in credit assessment.However,traditional feature selection methods are primarily based on linear correlations between features,making them ineffective at handling non-linear data relationships and accurately estimating the degree of correlation between variables.To overcome this challenge,this paper introduces an improved feature selection algorithm that combines the strengths of Random Forest and autoencoder.Firstly,irrelevant features with respect to the target variable are removed using Random Forest.Then,the collective importance of the remaining features is calculated,and these retained features are used to train an autoencoder neural network.Finally,the learned parameters from the autoencoder are utilized to initialize a three-layer neural network for feature reconstruction.Experimental validation was conducted on a publicly available credit assessment dataset,and the results indicate that the proposed algorithm outperforms other methods.

关 键 词：信用评估 随机森林 自编码器 特征选择 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]