摘要
广告点击率预估问题作为计算广告领域的核心问题,为了在数据量和用户信息急剧增多等情况下,自动挖掘特征之间的关系,增大重要特征的作用,提高模型的预测性能,提出一种新的融合结构——AP-XDeepFM模型。该模型在XDeepFM上进行改进,引入注意力机制,在DNN的第一个隐藏层前加入Product层,并将线性模型替换为FM模型。该模型可以更好地捕捉特征之间的关系,不仅能自动地进行显式的和隐式的高阶特征交互,而且能挖掘低阶特征。在avazu公开数据集上进行实验,结果表明改进的模型能有效提升模型的性能和泛化能力。
Click-through rate prediction is a key problem in the field of computing advertising. In order to automatically mine the relationship between features when the amount of data and user information increase rapidly, increase the role of important features, and improve the prediction performance of the model, this paper proposed a new fusion structure-AP-XDeepFM model. This model improved XDeepFM(eXtreme deep factorization machine) model, introduced attention mechanism, and introduced the product layer based on DNN(deep neural networks). In addition, the LR model was replaced by FM(factorization machines). This model could better extract the relationship between features. It could automatically combine explicit and implicit feature interactions, and create low order interactive features. The experimental results on avazu public data set show that the improved model can effectively improve the performance and generalization ability of the model.
作者
侯娜
邵新慧
Hou Na;Shao Xinhui(College of Science,Northeastern University,Shenyang 110819,Liaoning,China)
出处
《计算机应用与软件》
北大核心
2022年第12期108-113,131,共7页
Computer Applications and Software
