Research On Click-through Rate Prediction Model Based On User Preference Network

In the era of big data,mining and using mass user data to achieve more accurate and personalized recommendations and advertising has become a hot research spot in the industry.Click-through rate prediction plays an important role in advertising and recommendation,which directly affects the business revenue and user experience.Thus it has extremely important commercial value.User preference information hidden in the user behavior sequence plays an important role in the click-through rate prediction.Current research on click-through rate prediction mostly design a model from the perspective of modeling the interaction of low-order and high-order features,and it is difficult to fully mine and utilize the implicit user preference information hidden in the user historical behavior sequence.This paper reviews and summarizes the achievements of previous researches and explorations on click-through rate prediction.Based on this,we put forward two click-through rate prediction models based on user preference network from the perspective of modeling user preferences,include click-through rate prediction model based on users' short-term and long-term preferences and click-through rate prediction model incorporating user's drift preference.The mainly attributions in these two models include:(1)In view of the diversity,multi-dimension and local-activation of user preferences,a preference extraction network is proposed to mine users' short-term and long-term preferences.It divides the user preference into short-term preference and long-term preference learning respectively according to the characteristics of diversity and local-activation over time in user preferences,and design a preference extraction network learns the user's short-term and long-term preference from user behavior sequence.Multi-head attention mechanism is also designed to learn the fine-grained and multi-dimensional feature of users' long-term and short-term preference in different subspaces.(2)Aiming at the problem of user preference drift in real scenarios,a preference drift network based on PAGRU is proposed to model the drift of user preference.Inspired by the research in context-aware recommendations,PAGRU designed in this paper regards user's history preference information as a special kind of context information.Through the design of preference attention gate,the history preference information and user behavior sequence information are deeply integrated into the generation of forget gate,update gate and candidate state.(3)In view of the different influence of user preference features on click-throuth rate prediction,a preference fusion network based on SE attention mechanism is proposed to learn the complex interactive relationship between various features.By the design of squeeze and excitation structure,preference fusion network captures the cross-channel interaction relationship between multiple features.It learns the contribution of different preference features adaptively and various features are fused to get the preference fusion feature.Two international public datasets are used to verify the proposed click-through rate prediction models based on user preferences network.The final experiment shows that these two models proposed in our paper can fully exert the preference information hidden in user behavior sequence and obtain more effective user preferences features.Compared to other comparative click-throuth rate prediction models,the model proposed in this paper has an obvious improvement in AUC and Rela Impr.