基于改进XGBoost的金融客户投资行为特征选择方法  

Feature selection method for financial customer investment behavior based on improved XGBoost

作　　者：吴成英 马东方[2] WU Chengying;MA Dongfang(IT Department,Huachuang Securities Company Limited,Hangzhou Zhejiang 310051,China;Ocean College,Zhejiang University,Zhoushan Zhejiang 316021,China)

机构地区：[1]华创证券有限责任公司杭州研发中心,杭州310051 [2]浙江大学海洋学院,浙江舟山316021

出　　处：《计算机应用》2024年第S01期330-336,共7页journal of Computer Applications

基　　金：国家自然科学基金资助项目(52172334)。

摘　　要：金融客户投资购买行为是投资者动态购买理财产品交易决策的综合结果,受到客户自身属性、产品因素、行情信息和历史交易等多个不同因素的影响,原始因子属性的特征维度庞大、拟合风险偏高。现有研究主要通过不同的算法提高特征选择的准确率,忽略了不同群体的差异化特征及动态因素的影响。因此,提出一种改进XGBoost(eXtreme Gradient Boosting)的特征选择算法,并在金融客户投资行为上应用研究。针对客户群体投资行为的差异性,多维度综合量化分析投资行为,以解决单一投资行为指标不合理问题;对不同客户群体通过主成分分析(PCA)降维和优化的K-均值(K-means)聚类算法进行多属性融合聚类,然后分别对聚类后的不同群体使用改进XGBoost进行多分类预测,并通过修剪特征因子提升预测准确率。实验结果表明,使用改进XGBoost后,金融客户投资行为的特征因子维度更贴近实际,准确率更高。The investment and purchase behavior of financial customers is the comprehensive result of customer’s dynamic decision-making to purchase financial products,which is influenced by multiple different factors such as the customer’s own attributes,product factors,market information,and historical transactions.The original factor attribute has a large dimension and high fitting risk.Existing researches mainly improve the accuracy of feature selection by different algorithms,ignoring the differentiated features of different groups and the influence of dynamic factors.Therefore,an improved XGBoost(eXtreme Gradient Boosting)feature selection algorithm was proposed and applied to study on the investment behavior of financial customers.Aiming at the differences in the investment behavior of customer groups,a multi-dimensional comprehensive quantitative analysis was conducted to solve the unreasonable problem of a single investment behavior indicator.Multi-attribute fusion clustering was performed by PCA(Principal Component Analysis)downscaling and optimized K-means clustering algorithm for different customer groups,then the clustered groups were respectively predicted using improved XGBoost for multi-classification,and the prediction accuracy was improved by pruning the feature factors.Experimental results show that after using the improved XGBoost,the dimensions of the feature factors of the investment behavior of financial customers are closer to the reality and the accuracy is higher.

关 键 词：特征选择 XGBoost 多类别分类 主成分分析 K-MEANS聚类 投资行为 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]