Research On Personalized Recommendation Algorithm Based On Implicit Feedback

With the increasing visibility of Internet information overload,recommender systems have gained more and more attention because it can automatically find interesting information for users.Recommendation algorithm is the core of recommender systems,and the widespread attention to recommendation algorithm begins with the scoring prediction problem represented by the Netflix movie recommendation.With the changes in user habits and psychological expectations,entertainment platforms such as video,music,etc.need to recommend based on implicit feedback.Such problems are often more complicated than explicit feedback recommendations.At the same time,with the breakthrough of deep learning in the field of image processing and natural language processing,more scholars have begun to invest in the research of recommender algorithms based on deep learning,and have achieved certain development.This paper studies the personalized recommendation algorithm based on implicit feedback.A SE-WDL fusion model based on embedded sharing is proposed.Firstly,the user behavior records are extracted with statistical features of different dimensions.Then,for statistical features,user and item meta-attribute features,ID super-sparse features,and user behavior sequence combination features,the Wide module,Deep module,and LSTM module are designed separately.The Deep and LSTM embedded sharing method and joint training are adopted to achieve full integration of attribute features,statistical features and behavior sequence information.For the existing models,there is no discrimination between different types of implicit feedback behaviors of users,and some important types of feedback behavior data are sparse,which leads to low importance of personalized features and affects the recommendation of personalized performance.This paper focuses on the research based on feature embedding of user behavior types.Firstly,reasonable weight assignment method for different behavior types is designed,and feature embedding based on LFM is implemented.Then Attention Matching Net is further designed to introduce the"attention" mechanism into the user behavior sequence.The analysis module enables the model to automatically learn the weights of different behavior types,and the user images are fused into the user embedding vector so that more appropriate and accurate user and item feature embeddings are achieved.The embedded features generated by the two methods are used both in the collaborative filtering of the recall layer and the ranking model of the ranking layer,and have achieved good performance in the experimental data set of this paper.