Multifeedback Behavior-Based Interest Modeling Network for Adaptive Click-Through Rate Prediction

With the rapid development of Internet, recommendation system is becoming more and important in people’s life. Click-through rate prediction a crucial task system, which directly determines effect system. Recently, researchers have found that considering user behavior sequence can greatly improve accuracy click-through model. However, existing models usually use click as input model, will make it difficult for model to obtain comprehensive interest representation. In this paper, unified multitype modeling framework named MBIN, a.k.a. multifeedback behavior-based Interest network, proposed cope with uncertainties noisy data. The adaptive uses deep learning technology, obtains representation through multihead attention, denoises using vector projection method, fuses interests dropout technology. First, an denoising layer effectively mitigate noise problem sequences accurate interests. Second, fusion introduced so fuse various types representations users achieve personalized fusion. Then, we used auxiliary losses based on enhance effectiveness characterization. Finally, conduct extensive experiments real-world large-scale dataset validate our approach CTR tasks.