Research And Implementation Of The Video Recommendation Algorithm Based On Graph Model

The emergence and popularity of the Internet provide users a wealth of information and resources, and with the increasing for the demand of the information visualization, film-television and media, a lot of video resources begin to spread on the network. While the massive video resources meet the demand of quantity of person, it also brought difficulties for users to find the videos they are really interested, and finally, reduce their efficient use of video resources. In order to provide users the convenience to find the videos they wanted in a short time, a personalized video recommendation service coming into being, and it has been successfully applied in the major video sites, such as YouTube, Youku, iQiYi, etc.With the development of video sites, the users are usually able to conduct a variety of operations about the videos, such as collect, share and download the videos, and the users can also communicate and follow each other. The emergence of these multi-user behaviors and social behaviors provide new space and challenges to the existing algorithms of recommendation. To take full advantage of the two kinds of behaviors, we propose a graph algorithm of recommendation, which is based on multi-user behaviors and attention behaviors. It can calculate the proportion of each behavior accurately, and extract valid the relations of their attention through Bayesian model, then the random walk of algorithm begins on user behavior graph, and spread the access probability on the users attention graph. Finally composite the results of the two parts and put the highly accessed videos on their suggestion lists. In another word, calculate the personalized recommendations for users. The performance of the algorithm has been tested by Youku datasets and Epinions datasets this article crawled.In the actual scene, combine the multiple recommendation requests to form a recommendation request collection, and we called it multiple dimensions recommendation algorithm which provide recommendation service for recommendation request collection. By analyzing the distribution of the periods and types for people to watch videos, this thesis found two kinds of useful information for the multiple dimensions recommendation. In addition, since these video sites has accumulated massive user behaviors, some of these behaviors are no longer the representation of users’ interest, therefore, it is necessary to consider the attenuation of the proportion about user behaviors over time. Combining the two points above, this thesis presents a multiple dimensions recommendation algorithm based on the random walk and the time decay model. The algorithm calculates the relationship between users, videos and the information of multiple dimensions recommendation, then get the result of personalized recommendation through random walks. The performance of the algorithm was test on MovieLens datasets.Finally, the thesis also designed and implemented a personalized video recommendation system, which has a complete user functionality, video search capabilities, play online function, social function, and recommendation systems, and also put the algorithm this article proposed on the recommended system to provide users a better experience.