The Design And Implementation Of Personalized Movie Theme Playlist Recommendation System

Video recommendation is a hot topic in the field of recommendation system.Long videos are different from text.When recommending to users,they can show relatively little information.Users can only judge whether they are interested in video by name and poster most of the time.In order to solve these problems,this paper designs and implements a movie theme playlist recommendation system,which seeks solutions from the playlist point of view.On the one hand,it can help users find videos that they Are interested in and reduce the time of information screening.On the other hand,it can show movies to users who are interested in it,and improve the traffic of the platform,so as to achieve a win-win situation between users and platforms.Compared with the traditional recommendation system,the movie theme playlist recommendation system needs to solve the problems of theme finding and video aggregation.The system will mine the user's behavior,transform user log into abstract user persona,automate the production of interest playlists,content-based playlists and non-personalized playlists,and then through multi-channel recall,mixed recommendation,diversity filtering and other steps,recommend the playlists precisely to the users who need them.The system is divided into five modules,which are user portrait module,broadcast order processing module,recommended engineering module,front-end interface module and engineering monitoring module.The calculation of user portrait mainly uses the play index formula,and designs the weight calculation formula based on TF-IDF(Term Frequency-Inverse Document Frequency)and the memory forgetting curve incremental model.In the automatic expansion of the broadcast order,it is necessary to calculate the co-occurrence similarity between the movie and the movie.In the case of multiple recalls,the cosine similarity of the feature vector of the user image and the broadcast tag vector is mainly used,and based on content and non-personalized multi-channel recall.When recommended,it is mainly designed as a control group,weighted SlopeOne algorithm,ALS(Alternating Least Squares)matrix decomposition,before the output of the broadcast,the recommendation list needs to be stack-based Diversity filtering to ensure that movies in the returmed order are not over-repeated.The front-end interface module uses the Spring Boot framework to provide interface services to clients.The project mainly uses Spark as the calculation engine,Spark MLlib machine learning library,Hive and Hbase as big data storage tools,and Couchbase as an online database to implement a complete recommendation system.After the project was launched,bucket test was used to monitor and verify the recommendation effect online.Compared with the old interface,the CTR(Click-Through-Rate)and UCTR(User Click-Through-Rate)increased significantly.The project uses a wide range of scenarios,and has been applied in TV and mobile multi-channel,which can have different recommendation effects for different display blocks.