基于多路交叉的用户金融行为预测  

Prediction of user financial behavior based on multi-way crossing

作　　者：程鹏超 杜军平[1] 薛哲 CHENG Pengchao;DU Junping;XUE Zhe(Beijing Key Laboratory of Intelligent Telecommunication Software and Multimedia,School of Computer Science,Beijing Uni-versity of Posts and Telecommunications,Beijing 100876,China)

机构地区：[1]北京邮电大学智能通信软件与多媒体北京市重点实验室,北京100876

出　　处：《智能系统学报》2021年第2期378-384,共7页CAAI Transactions on Intelligent Systems

基　　金：国家重点研发计划项目(2018YFB1402600);国家自然科学基金项目(61772083,61802028);广西科技重大专项(桂科AA18118054).

摘　　要：针对通过挖掘用户的金融行为来改善金融领域的服务模式和服务质量的问题,本文提出了一种基于多路交叉特征的用户金融行为预测算法。根据数据包含的属性构建训练的特征,基于因子分解机模型(FM)利用下游行为预测任务对金融数据的特征进行预训练,获取数据特征的隐含向量。引入特征交叉层对金融数据的高阶特征进行提取,解决FM线性模型只能提取低阶特征的缺点。利用残差网络对金融数据的高阶特征进行提取,解决深度神经网络在提取金融数据高阶特征时由于网络层数过深而导致的梯度消失的问题。最后,将FM、特征交叉网络和残差网络整合为统一的多塔模型进行用户金融行为预测,并融合低阶特征与高阶特征进行用户金融行为预测。在多个数据集上对算法的有效性进行了实验验证,实验结果表明,所提出的算法能够取得较好的用户金融行为预测的准确率。To improve the service mode and service quality in the financial field by mining the financial behaviors of users,a user financial behavior prediction algorithm based on multi-way crossing(MCUP)is proposed in this paper.First,the training features are constructed based on the attributes contained in the data.Based on the FM model,the downstream behavior prediction tasks are used to pre-train the features of the financial data,and the hidden vectors of the features are obtained.Second,the feature cross-layer is introduced to extract high-order features of financial data,overcoming the disadvantage that the FM linear model can only extract low-order features.Then,the residual network structure is used to extract high-order features of financial data,solving the gradient disappearance problem caused by the too deep network layer.Finally,a unified multi-tower model integrated by the FM,feature cross network,and residual network is used to predict user financial behavior,blending low-order and high-order features.Experimental results show that the proposed algorithm can achieve a better accuracy rate in predicting user financial behavior.

关 键 词：行为预测 金融 多路交叉 残差 多塔模型 预训练 挖掘 联合训练 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]