基于域矩阵因子分解机的点击通过率预估增强网络  

Enhanced network for CTR prediction based on field-matrixed factorization machines

作　　者：陈乔松 黄泽锰[1,2] 胡静 王进 邓欣[1,2] CHEN Qiaosong;HUANG Zemeng;HU Jing;WANG Jin;DENG Xin(School of Computer Science and Technology,Chongqing University of Posts and Telecommunications,Chongqing 400065,P.R.China;Key Laboratory of Data Engineering and Visual Computing,Chongqing University ofPosts and Telecommunicatioins,Chongqing 400065,P.R.China)

机构地区：[1]重庆邮电大学计算机科学与技术学院,重庆400065 [2]重庆邮电大学数据工程与可视计算重点实验室,重庆400065

出　　处：《重庆邮电大学学报（自然科学版）》2024年第2期383-392,共10页Journal of Chongqing University of Posts and Telecommunications(Natural Science Edition)

基　　金：国家重点研发项目(2022YFE0101000)。

摘　　要：有效的特征交互,对于工业推荐系统中点击通过率(click-through-rate,CTR)预估的准确性起着至关重要的作用。以往并行结构的CTR预估模型通过将独立的浅层模型和深层模型并行连接,以此来学习特征的低阶交互和高阶交互。但是,这些模型存在浅层模型准确性低、未考虑特征交互时的多语义问题、参数过多、深层模型过度泛化等问题。基于上述问题,提出了一种基于域矩阵因子分解机的点击通过率预估增强网络,通过引入域矩阵优化浅层模型中的交互,提高运算效率,并在深层模型的DNN层与层之间增加了桥接模块,在每层高阶交互后增强对原始特征的记忆能力,将浅层模型和深层模型的结果相加并归一化得到预测值。该模型在Criteo、KKBox、Frappe和MovieLens数据集上进行了大量实验,展现了优秀的预测能力。Effective feature interaction plays a vital role in the accuracy of click-through-rate(CTR)estimation in industrial recommendation systems.Previous CTR prediction models with a parallel structure learn low-order and high-order interactions of features by connecting independent shallow models and deep models in parallel.However,these models have problems such as low accuracy of shallow models,failure to consider the multi-semantic problem of feature interaction,excessive parameters,and over-generalization of deep models.Based on the above problems,this paper proposes an enhanced network for CTR prediction based on field-matrixed factorization machines.It introduces domain matrix to optimize the interaction in shallow models,improves the efficiency of computation,and adds a bridge module between the DNN layers of deep models to enhance the memory ability of original features after each high-order interaction.The results of shallow and deep models are added and normalized to obtain the predicted value.The model has undergone extensive experiments on Criteo,KKBox,Frappe,and MovieLens datasets,demonstrating excellent predictive capabilities.

关 键 词：点击通过率 域矩阵因子分解机 桥接模块 特征交互 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP391[自动化与计算机技术—控制科学与工程]