Research On Click-through Rate Prediction Problem In Recommender Systems

With the rapid development of the Internet and the explosion of data,recommendation systems have been widely used due to the needs of both network service providers and users.A good recommendation system can not only allow users to effectively obtain the services they need and improve user satisfaction,but also allow service providers to attract users and generate commercial value that cannot be ignored.For recommendation systems,the problem of click-through rate prediction is crucial.With the continuous development of machine learning in recent years,click-through rate prediction models based on machine learning methods are now widely used,and these models have achieved great success,but there are still some problems.They perform simple feature interaction,such as inner product operation,without distinguishing the different importance of different features.The click-through rate prediction model is usually offline and rarely uses online user feedback for learning,and there is less research on the cold start problem.This thesis studies the above problems in click-through rate prediction.Specifically,the main contents of the thesis include the following two aspects:(1)Many current click-through rate prediction models just compute feature interactions in a simple way,but they are less concerned with the different importance and computational requirements of different feature interactions.The general idea is that some complex feature interactions may require more computation to produce the final result,while some simple or unimportant feature interactions may require less computation.Inspired by this idea,the thesis proposes an adaptive deep attention model,a new model that automatically learns the interaction of high-order features from raw data.The core of the model is a multi-head self-attention neural network that learns feature interactions and a network depth control module that controls the network depth required for different feature field interactions.This thesis has conducted extensive experiments and the results show that the proposed model has better predictive performance compared to other models.(2)For the problem that the current model rarely uses the real-time online feedback data of users,the thesis studies the problem of click-through rate prediction based on reinforcement learning.The interaction mode of the reinforcement learning environment and the agent are naturally suitable for the online click-through rate.The prediction model can efficiently use the user's online feedback data to track the user's interest changes,and optimize the model in real time.In the thesis,a reinforcement learning environment is constructed using the real data set,and a click-through rate prediction model based on the deep Q-Network is proposed.At the same time,a novel reward mechanism is proposed,which verifies that the proposed model has a good click-through rate prediction effect in the case of non-cold start and user cold start.