Session interest model for CTR prediction based on feature co-action network

The main purpose of click-prediction models is to predict the probability of customers clicking on products and provide support for advertising decisions of businesses. However, most previous models often use deep neural networks to capture implicit interaction and can not fully retain the representational power of the original feature interactions. At the same time, one factor that most models ignore is that sequence is made up of sessions. Therefore, how to model user interest features and preserve the representational properties of feature interactions is the main challenge to improve the accuracy of CTR prediction. According to above issues, this study propose session interest model with feature co-action network (SIFAN). First, we used widely used characteristic co-action network module to tap into the interactions in customer single behavior. Then, the sequential behavior of customers is divided into session layers, and considering that various session interests may follow sequential patterns, gated recursive units are applied to predict customer clicks. Then, by analyzing the GRU with attention update gates, the correlation between conversation interest and target items is determined. According to relevant experimental results, under the same experimental conditions, the SIFAN model has significant performance advantages compared to other models.