Research On Click Through Rate Prediction Based On Graph Neural Network

Click Through Rate prediction is an important task of commercial recommendation system,the key of which is feature interaction.Accurate prediction of the probability of users' clicking on advertisements enables advertising platforms to generate great economic benefits,enables advertisers to gain short-term and long-term benefits,and improves user experience.However,modeling effective feature interactions is challenging because feature interactions are extremely complex and flexible and the advertisements in click-through data of real scenarios have cold start problem.In this paper,the problem of click-through rate prediction is analyzed and studied.Combined with the feature information and field information in the data,two graph neural network models are proposed to predict the click-through rate.In the first stage,in order to model the influence of field on feature interaction explicitly,a Field-Aware Feature Interaction via Graph Neural Network(called FAFI-GNN)is proposed.Firstly,the model gets feature representation and field representation through the presentation layer.Next,the model regards features as the node and builds feature graph.In order to model the influence of field on feature interaction,a field-aware adjacency matrix calculation module is designed to calculate the interaction intensity of feature nodes according to the field information.Then,the high-order features are obtained through multiple iterations of the gated graph neural network.Finally,effective cross features are selected through the attention mechanism for fusion,and the fusion results are input into the prediction layer for final prediction.In the second stage,aiming at the cold start problem in the data of real business scenarios,multi-granularity interactions are constructed to enrich the information and improve the prediction accuracy.Based on FAFI-GNN,a Dynamic Merge of MultiGranularity Feature via Graph Neural Network(called DMGF-GNN)is proposed.FAFIGNN models the interaction at feature level,which is fine-grained feature.On this basis,DMGF-GNN builds a coarse-grained feature extraction module,and extracts coarsegrained features from field information through graph neural network to supplement effective information and improve model performance.In addition,in order to capture the user's preference for field information,a cross-reference module is designed.The finegrained feature graph and coarse-grained field graph cross-reference each other's information to calculate the adjacency matrix,which not only explicitly captures the influence of field on feature interaction,but also realizes the personalized modeling of coarsegrained feature.Finally,a fusion layer based on attention mechanism is constructed to dynamically fuse the features of different granularity to predict the click-through rate.In this paper,the FAFI-GNN model and DMGF-GNN model are evaluated on three public datasets,and the experimental results proved the effectiveness of the two models.And the performance of AUC was significantly better than other existing methods.DMGF-GNN model integrates coarse-grained and fine-grained features,so captures more abundant information,and its prediction accuracy is higher than that of FAFI-GNN model.