Research On Collaborative Recommendation Algorithm Based On Deep Learning For Video Program

At present,with the rapid advancement of the intelligent platforms,users can download or watch the network TV programs more easily.However,how to accurately express users' preferences and recommend the video programs they want from the massive data is still a problem that has not been solved.For businesses,how to recommend suitable TV programs for different users plays an important role in improving program ratings and purchase rates.At present,most of the researches only use the rating information of users,it can't settle the disputes in the above recommendation.In the face of massive film and television programs,users can only access part of the program information,resulting in sparse historical record data that can reflect users' preferences.With the development of network,we can obtain more and more auxiliary information(such as images and text)in different channels,so how to effectively integrate these multi-source auxiliary information in recommendation systems for different purposes,making accurate recommendation by users has become an important research issue in the field of recommendation systemsThis thesis aims to solve the problem of multi-source auxiliary information fusion recommendation in the process of films and television programs recommendation.The main work and technological innovation research focus are as follows(1)This thesis introduces a way of automatic feature extraction to enhance the traditional matrix decomposition model.In the first mock exam,the traditional matrix decomposition model is improved.The users and films features are modeled by convolution neural network,and the potential factors of users interest are modeled.On this basis,the interest factor feature matrix containing relevant information output by the neural network has replaced the traditional feature matrix,and the feature vectors of both users and film and television programs are inner product Processing,get the corresponding results.After conducting experiments,we conducted a lot of analysis on the results.Compared with the traditional model,this algorithm can effectively improve the predictive ability of the score data,and the accuracy of prediction is also larger than that of the traditional algorithm.(2)In this thesis,we propose a compound recommendation model that combines multiple sources of auxiliary information.Because the matrix decomposition model essentially uses the linear relationship to represent the relevance with users and items,which has certain limitations,this thesis proposes a deep composite recommendation model based on the variational autoencoders,which uses the deep learning technology to model the non-linear users and items relationships from the rich accessible data sources,so as to accurately reflect the users preferences.Based on the traditional variational autoencoders,the original simple Gauss priori is improved,and the heterogeneous multi-source auxiliary information features are taken as a new priori distribution to enrich the data space.In addition,since the independent variational autoencoders can not generate the data distribution well after incorporating the auxiliary information,this thesis introduces the generative adversarial networks to improve the reconstruction target of the variational autoencoders,the improved network structure can be better applied to the recommendation systems with sparse data.Finally,this thesis collects a large-scale data set from a movie website,which contains the user's comments and posters of the films and TV programs.A large number of experiments on this data set reveal that the algorithm has significantly improved the Recall and NDCG indicators compared with the most advanced benchmark method in films and video programs recommendation.