广告点击率预估的逐层残差交互网络  

Layer-by-Layer Residual Interactive Network Approach for Advertisement Click-Through Rate Prediction

作　　者：尹云飞[1,2] 龙连杰 黄发良 吴开贵 YIN Yun-Fei;LONG Lian-Jie;HUANG Fa-Liang;WU Kai-Gui(College of Computer Science,Chongqing University,Chongqing400044;Guangxi Key Lab of Human-Machine Interaction and Intelligent Decision,Nanning Normal University,Nanning530100)

机构地区：[1]重庆大学计算机学院,重庆400044 [2]广西人机交互与智能决策重点实验室,南宁530100

出　　处：《计算机学报》2024年第3期575-588,共14页Chinese Journal of Computers

基　　金：国家自然科学基金(61962038);中央高校基本科研业务费项目(2022-CDJKYJH023);广西人机交互与智能决策重点实验室开放基金项目(GXHIID2208)资助.

摘　　要：网络广告费的收取通常是以用户的点击次数来计算的,因此如何准确地预估点击率(CTR)是广告公司十分关心的问题.当前先进水平的方法集中在构建各种高阶特征交互模型来预估CTR,但是高阶特征交互会丢失低阶信息,尤其是丢失原始特征的信息.为此,本文提出一个新的逐层残差交互网络,它在每次交互时都考虑原始特征的引导作用,被命名为逐层残差交互网(LRIN).LRIN强调高阶特征交互应该建立在原始特征逐层交互的基础上.n阶特征交互由原始特征与n-1阶特征通过元素积运算得到.进而,本文引入了多尺度方法来设计注意力网络.受逐层交互的影响,注意力网络也被设计成多层,称之为逐层注意力网络.为了将二者结合起来,本文提出将逐层残差交互网络的输出作为逐层注意力网络的权重,由此形成了一种新的双网络训练模型.在多个benchmark数据集上的实验结果表明,LRIN的性能比当前先进的方法在Criteo数据集上平均提高1.24%,在Avazu数据集上平均提高2.16%,在MovieLens-1M数据集上平均提高了1.3%,在Book-Crossing数据集上平均提高了1.27%.Online advertising fees are charged based on the number of times that users click on ads,and therefore how to accurately predict Click-Through Rate(CTR)is a very concerned issue for advertising companies.Current state-of-the-art methods focus on constructing various high-order feature interaction models to predict CTR;however,high-order feature interactions will lose low-order information,especially the information of original features.To this end,a novel layer-by-layer residual interaction network framework is proposed in this paper,which leverages the guiding role of the original features at each interaction,and is named as the Layer-by-layer Residual Interaction Network(LRIN).LRIN emphasizes that higher-order feature interactions should be based on the interactions of original features layer by layer.The interaction of n-order features is obtained by the element-wise product between the original features and the n-1-order features.Moreover,a multi-scale approach is introduced to design attention network.Affected by layer-by-layer interaction,the attention network is also designed into multiple layers,which is called layer-by-layer attention networks.In order to combine the two,this paper proposes to take the outputs of the layer-by-layer residual interaction network as the weights of the layer-by-layer attention network,and thus forms a novel dual-network training model.The experimental results on multiple benchmark datasets indicate that the performance of LRIN is on average 1.24%better than current advanced methods on the Criteo dataset,2.16%better on the Avazu dataset,1.3%better on the MovieLens-1M dataset,and 1.27%better on the Book-crossing dataset.

关 键 词：残差网络 逐层 特征交互 CTR预估 注意力 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]