基于特征与域感知的点击率预估方法

点击率预估是推荐系统中的核心任务，其关键是学习有效的特征交互，但现有基于深度神经网络的点击率预估方法未考虑冷启动问题，导致准确率降低。结合特征信息和域信息的嵌入，提出一种特征交互的点击率预估方法FF-GNN。利用基于图神经网络的交互模块分别提取特征嵌入和域嵌入的结构信息，建模细粒度的特征交互和粗粒度的域交互过程。同时通过设计图神经网络的权重计算模块，交叉引用特征图神经网络和域图神经网络的低阶特征信息，实现特征交互和个性化建模域交互。在此基础上，采用注意力机制融合特征交互和域交互模块的结果预测点击率。在Criteo和Frappe公开数据集上的实验结果验证了FF-GNN方法的有效性，其AUC指标相较于同类型Fi-GNN方法分别提高0.57和0.85个百分点，能够同时关注特征和域信息，提高点击率预估的准确度。

CTR Prediction Method Based on Feature and Domain Perception

CTR prediction,as the core task in the field of recommendation systems,is key to learning effective feature interaction. The existing CTR prediction method based on deep neural network do not consider the cold start problem,resulting in the low accuracy. For feature interaction,this paper proposes the CTR prediction method FF-GNN which combines the embeddings of feature and domain information. The interaction module based on Graph Neural Network(GNN)is used to extract the structural information of feature and domain embeddings to model the fine-grained feature interaction as well as the coarse-grained domain interaction.In addition,a weight calculation module is designed for the GNN. This module cross references the low-order feature information of the feature GNN and domain GNN to realize feature interaction and personalized modeling domain interaction. The attention mechanism is then used to fuse the results of feature interaction module and domain interaction module to predict CTR. Experimental results on the Criteo and Frappe datasets verify the effectiveness of the FF-GNN method.Compared with the Fi-GNN method,the AUC of FFGNN is improved by 0.57 and 0.85 percentage points respectively. The method by focusing on the feature and domain information,improves the accuracy of CTR prediction.