Click-Through Rate Prediction Based on User Behaviors and Feature Importance

Abstract Click-through rate prediction is one of the hot topics in the field ofrecommendation and advertising systems. The existing CTR predictionmodels mainly include feature interaction and behavior sequence. Forexample, PNN, DCN and other models use deep learning methods toextract feature interaction information, while DIN and DIEN captureuser interest through user’s historical behavior, and then use attentionmechanism to model the relationship between target items and behavior sequences. However, the existing CTR prediction techniques eitherignore both aspects or only consider one of them, which limit the prediction performance. In order to solve the above problems, we proposea click-through prediction model CUBFI that combines user behaviorand feature importance in this paper. Firstly, In order to extract interests, Global-local Gate and Post-LN Informer module are proposed atthe interest extraction layer. In addition, we introduce auxiliary losses tosupervise the extraction of user interest features. Secondly, In the interestupdate layer, we introduce A-GRU to enhance the relationship betweeninterest expression and target items. Finally, for non-temporal features,this paper proposes a multi-cross layer to increase the nonlinear abilityof the model. Experiments show that our model can effectively improve the click-through rate prediction accuracy of advertisements. The codes will be available at https://github.com/jihuiqin2/sequence ctr.