Research And Application Of Deep Recommendation Model Based On Attention Mechanism

In order to solve the problem of information overload in the Internet,recommendation system has been widely used in various applications,such as video websites,e-commerce platforms,etc.From the development of digital TV,IPTV enables audience to access multimedia content in a more convenient and efficient way.In particular,IPTV provides interactive services that enable audience to view preferred content at any time.However,with the continuous increase of available content,IPTV also faces the dilemma of "difficult user selection".Therefore,IPTV operators develop recommendation systems to provide personalized services to the audience.This thesis focuses on the recommendation model for IPTV applications,and specifically uses the attention mechanism in deep learning to build the corresponding recommendation model.Specifically,for the implicit feedback problem in IPTV applications,the model Transformer with Fusion is proposed to model the behavior sequence of users;for the problem that IPTV accounts are shared by family users,the model Trans-LSTP is proposed to model the behavior sequence of users.The thesis mainly completed the following three aspects of work.1.According to the user's sequence behavior,this paper has proposed Transformer with Fusion based on Transformer and MLP to solve the problem that there is no explicit feedback in IPTV application,but abundant implicit feedback.Specifically,the semantic layer based on the transformer models the user's historical sequential behavior,and decodes the target movie to obtain the user's semantic preference;the fusion layer captures the linear and nonlinear correlation between the user's semantic preference and the target movie.This method has been verified in IPTV application online recommendation system.At the same time,the experimental results of Movielens dataset and Amazon Product dataset also prove the effectiveness of this method.2.This thesis has proposed a model named Trans-LSTP to solve the problem that IPTV accounts are usually shared by a group of family users,and the preferences of each family member may be different.In order to understand the preferences of home users more dynamically,the model introduces the long-term preferences of home users and the short-term interests of current users.Although the preferences of each family member are different,this thesis believes that each member will be affected by the family and choose films and TV works that meet the public taste.The attention mechanism in Transformer automatically assigns the weight of items to dynamically capture the long-term and short-term interests.Meanwhile,the hierarchical design adaptively combines the long-term and short-term preferences and pays attention to their different effects on each member.The experimental results of Amazon dataset and subcategory dataset also show that the performance of this method exceeds all the baseline models.3.In order to run the model in the online recommendation system of IPTV application,this thesis also studies the practice of the deep recommendation model in the recommendation system,including data preprocessing and feature engineering,model training and storage,model deployment and inference,which are important parts of the industrial recommendation system.