刘弘历, 武森, 魏桂英, 李新, 高晓楠. 基于深度神经网络的点击率预测模型[J]. 工程科学学报, 2022, 44(11): 1917-1925. DOI: 10.13374/j.issn2095-9389.2021.03.23.002
引用本文: 刘弘历, 武森, 魏桂英, 李新, 高晓楠. 基于深度神经网络的点击率预测模型[J]. 工程科学学报, 2022, 44(11): 1917-1925. DOI: 10.13374/j.issn2095-9389.2021.03.23.002
LIU Hong-li, WU Sen, WEI Gui-ying, LI Xin, GAO Xiao-nan. Click-through rate prediction model based on a deep neural network[J]. Chinese Journal of Engineering, 2022, 44(11): 1917-1925. DOI: 10.13374/j.issn2095-9389.2021.03.23.002
Citation: LIU Hong-li, WU Sen, WEI Gui-ying, LI Xin, GAO Xiao-nan. Click-through rate prediction model based on a deep neural network[J]. Chinese Journal of Engineering, 2022, 44(11): 1917-1925. DOI: 10.13374/j.issn2095-9389.2021.03.23.002

基于深度神经网络的点击率预测模型

Click-through rate prediction model based on a deep neural network

  • 摘要: 针对现有深度神经网络点击率预测模型在对用户偏好建模时,难以有效且高效地处理用户行为序列的问题,提出长短期兴趣网络(Long and short term interests network, LSTIN)模型,充分利用用户历史记录上下文信息和顺序信息,提升点击率预测精准性和训练效率。使用基于注意力机制的Transformer和激活单元结构完成用户长、短期兴趣建模,对用户短期兴趣进一步使用循环神经网络(Recurrent neural network, RNN)、卷积神经网络(Convolutional neural networks, CNN)进行处理,最后使用全连接神经网络进行预测。在亚马逊公开数据集上开展实验,将提出的模型与基于分解机的神经网络(DeepFM)、深度兴趣网络(Deep interest network, DIN)等点击率预测模型对比,结果表明提出的模型实现了考虑上下文信息和顺序信息的用户历史记录建模,接受者操作特征曲线下面积(Area under curve, AUC)指标为85.831%,相比于基础模型(BaseModel)提升1.154%,相比于DIN提升0.476%。且因区分用户长、短期兴趣,模型能够在提升预测精准性的同时保障训练效率。

     

    Abstract: The click-through rate (CTR) prediction task is to estimate the probability that a user will click an item according to the features of user, item, and contexts. At present, CTR prediction has become a common and indispensable task in the field of e-commerce. Higher accuracy of CTR prediction results conduces to present more accurate and personalized results for recommendation systems and search engines to increase users’ actual CTR of items and bring more economic benefits. More researchers used a deep neural network (DNN) to solve the CTR prediction problem under the background of big data technology in recent years. However, there are a few models that can process time series data and fully consider the context information of users’ history effectively and efficiently. CTR prediction models based on a DNN learn users’ interests from their history; however, most of the existing models regard user interest, ignoring the differences between the long-term and short-term interests. This paper proposes a CTR prediction model named Long- and Short-Term Interest Network (LSTIN) to fully use the context information and order information of user history records. This use will help improve the accuracy and training efficiency of the CTR prediction model. Based on the attention mechanism, the transformer and activation unit structure are used to model long-term and short-term user interests. The latter is processed using the recurrent and convolutional neural networks further. Eventually, a fully-connected neural network is applied for prediction. Different from DeepFM and Deep Interest Network (DIN) in experiments on an Amazon public dataset, LSTIN achieves modeling with context and order information of user history. The AUC of LSTIN is 85.831%, which is 1.154% higher than that of BaseModel and 0.476% higher than that of DIN. Besides, LSTIN achieves distinguishing the long-term and short-term interests of users, which improves the performance and maintains the training efficiency of the CTR prediction model.

     

/

返回文章
返回