| With the widespread use of mobile devices and the massive growth of video website resources,mobile viewing is becoming more and more popular.Watching movies on mobile devices has the advantages of being less affected by time and place,and low cost,but the explosive growth of movie resources make it difficult for users to quickly make effective and accurate movie options.Therefore,establishing a personalized video recommendation system to give users high-quality movie viewing services is the key to improving user satisfaction and website revenue.For the sake of those problems,this dissertation studies the user’s mobile viewing behavior based on user access records provided by the operator and movie content information from network and achieve the following innovative results.A rating category prediction method based on objective information of the film is proposed.The knowledge graph and natural language processing methods are used to process the structured and brief information of the film,and the deep neural network is used to automatically learn the information vector features to build a rating category prediction model.Achieved 62.1% accuracy and 63.9% F1-micro value on the movie rating three-class prediction problem,better performance than unimproved Trans R method and single information source model.Finally,the model successfully achieved the "new movie" rating prediction,effectively solving the "cold start" problem of new movie ratings.A hybrid recommendation system combining user viewing records and movie content information is proposed.Based on the traditional collaborative filtering,this model adds auxiliary information such as film structured information,film profile information,and film review information,and uses the attention mechanism to process different user comments.The above information features are automatically learned by the neural network without any manual feature technology.On the user’s mobile viewing movie dataset,the model’s HR @ 10 value reaches 0.564,the MRR reaches 0.307,and the recommended video coverage reaches 0.665,the performance is better than 6 traditional recommendation methods and models.Experimental results show that combining user viewing records and movie content information effectively improves movie recommendation performance and alleviates the problem of "long tail" movie recommendation. |
