基于分层注意力图神经网络的点击率预测模型

Model based on hierarchical attention and graph neural network for CTR prediction

点击率预测是推荐系统和在线广告中的一项基本任务,大多主流模型主要通过高阶特征和低阶特征交互建模以提高模型性能和泛化能力,然而很多模型只学习了每个特征的固定表示而没有考虑在不同上下文中每个特征的重要性。针对基线模型(Feature Refinement Network,FRNet)在不同上下文无法灵活处理重要特征选择,并且缺乏良好解释性的问题,提出了一种特征细化分层注意力图神经网络(Feature Refinement Graph Neural Network and Hierarchical Attention,FRGNN-HA)模型。首先,在基线模型中融合图神经网络结构,利用图神经网络聚合邻节点和自身节点特征以实现在非欧式空间新节点的表示向量的更新,从而提升在不同上下文的重要特征选择能力和良好的解释性。其次,在图神经网络的基础上设计分层注意力网络,让模型可以更好地自适应关注重要上下文信息,并且可以在噪声和复杂场景下依然保持较好的性能。最后,FRGNN-HA通过在Criteo、Frappe和MovieLens这3个数据集上对比实验结果表明,与基线FRNet模型相比,曲线下的面积(Area Under Curve,AUC,记为AUC)指标分别提升了0.07%、0.29%和0.06%,交叉熵损失函数Logloss(记为Lloss)分别降低了0.08%、0.81%和1.09%。

Click-through rate prediction is a basic task in recommender systems and online advertising.Most mainstream models mainly improve model performance and generalization ability by modeling high-order feature interactions and loworder feature interactions.However,many models only learn the fixed representation of each feature without considering the importance of each feature in different contexts.Aiming at the problem that Feature Refinement Network(FRNet)models cannot flexibly process important feature selection in different contexts and lack good interpretation,a Feature Refinement Graph Neural Network and Hierarchical Attention(FRGNN-HA)model is proposed.Firstly,the graph neural network structure is integrated into the baseline model,and the graph neural network is used to aggregate the features of adjacent nodes and their own nodes to update the representation vector of new nodes in non-European space,so as to improve the important feature selection ability and good interpretability in different contexts.Secondly,based on the graph neural network,a hierarchical attention network is designed,so that the model can better adaptively focus on important context information and can still maintain good performance in noise and complex scenes.Finally,FRGNN-HA compares the experimental results on the three datasets of Criteo,Frappe and MovieLens.The results show that compared with the baseline FRNet model,the Area Under Curve(AUC)index is increased by 0.07%,0.29%and 0.06%,and the cross entropy loss function Logloss is reduced by 0.08%,0.81%and 1.09%,respectively.