Application Research Of Topic Model In Mobile Video Personalized Recommendation Service

With the rapid development of the wireless network technology and the popularity of smart phones, tablet and other portable network terminal, more and more users choose to use mobile devices to obtain Internet services, especially video services. With the influence of massive video content and mobile device screen size, battery life, network traffic and other factors, mobile users are confronted with more serious information overload. The recommended system, as an effective way to alleviate the phenomenon of information overload, has been brought more and more attention.For cold start problem of mobile video Recommended system( including the new video and the Nagao video which play times is relatively low), this article focuses application of word vector topic model based on the Deep Neural Network(DNN) in mobile video personalization push notification service, we propose a mobile video recommended algorithm based on DNN. Firstly, for the given video text information(profiles, tags, description, etc.), this article obtain video distribution vector presentation by using vector expression model based on word vector; then constructing positive and negative behavior distribution vector of users according to the behavior of clicking on the video or not; finally generating the recommendation by comparing video vector feature deviation between positive and negative behavior distribution vector of users. With offline and online experiments in a large-scale mobile online video system, we evaluate our proposal and compare it with the random method, regional push method, contentKNN algorithm and ItemCF algorithm. The experiment results show that our proposal's performance has improved by as much as 106%, 88%, 41% and 57% respectively on the clickthrough rate in mobile online video services.Considering the traditional interest-based modeling based on vector form and the DNN algorithm usually only build a single user interest vector, which is not conducive to fine recommendation. This paper proposes a method of user multiple classifiers interest model, which establish multi categories of interest-vector according different message types and expand user interest modeling behavior data(including users of different sources of the video click play, offline download, concerns, comments and other actions),as a result, the number of users increased. The large scale experiments also show that the improved DNN algorithm has a certain improvement than the original recommendation effect.