Personalized Recommendation System For Mobile Videos

In recent years,with the rapid development of cloud computing,big data,mobile Internet,as well as the endless variety of video services applications have triggered explosive growth in mobile video data.Users can view and share their interested mobile video anytime and anywhere as well as constantly create and upload a wide variety of mobile videos.Due to the characteristic of being wide variety,single content,and short duration,mobile(short)video has become the main way of entertainment for people to spend leisure time in modern life.How to quickly and effectively find the interested videos from large-scale mobile videos and to provide a personalized recommendation service has become an inevitable trend in the development of mobile video browsing applications.In view of diversified mobile videos,it has practical significance for improving the performance of mobile video personalized recommendation system how to fully exploit user preferences from large-scale mobile videos.When users share and distribute mobile videos,they usually add a text tag to explain the video content.As an important subjective semantic clue of users,tag not only indicates the latent semantic information of the mobile video,but also reflects the user's personalized preferences.Therefore,effective tag information can contribute to user preference model construction and personalized recommendation.More recent,deep learning as the latest research achievement in the field of artificial intelligence has shown a new idea for personalized recommendation of mobile video.Accordingly,it is a practical engineering application problem for personalized recommendation technology of mobile video how to explore deep learning technology to extract deep features from mobile videos and construct user preference model with tags.Based on the above considerations,a personalized recommendation system for mobile video is designed and implemented in this thesis.The mainly works are as follows:(1)Deep features extraction of mobile videos based on ELU-3DCNNFirstly,the traditional three-dimensional convolutional neural network(3DCNN)is improved by using exponential linear unit(ELU)as the activation function.Then,the trained ELU-3DCNN network is utilized to extract deep features of mobile video after optimizing the ELU-3DCNN network with gradient-descent calculation.Finally,the extracted deep features are applied into the video classification data set to evaluate the performance of the proposed method.Experimental results show that the deep features of mobile video with our method can obtain higher classification accuracy over the state-of-arts feature representation methods.(2)User preference model construction with joint deep features-tagsFirstly,user tag preference model is constructed with the behavior data for users to browse video and mobile videos' s tag by using moving average method.Then,user deep feature preference model is constructed with behavior data for users to browse video and deep features of mobile video by using maximum likelihood estimation.Finally,the user preference model with joint deep features-tags is constructed with weighted voting method.The experimental results show that the constructed user preference model can accurately describe the user's preferences.(3)Personalized recommendation system for mobile video based on user preference modelFirstly,the user preferences are learned by using user preference model with joint deep features-tags.Then,the key frames of mobile video are detected by using differential evolution algorithm and GoogLeNet deep network to generate a summary description for recommended video.Finally,based on the user preference model,a personalized recommendation system for mobile video is implemented.The experimental results show that the proposed personalized recommendation system can improve the recommendation accuracy of mobile video.