A Study Of Movie Recommendation System Based On The Implicit Social Networks

ABSTRACT:As the development of information technology and the popularization of network, online video service spreads rapidly, making it become one of the most traffic contributors on the Internet. The emerging video systems provide various wonderful programs which enrich people’s life. However, it is the abundant resources that make choice difficult. Traditional personal recommendation recommends videos according to users’interest. However, when users expand, it will be challenging to maintain the speed of data transmission and the stability of system. On the other hand, group recommendation recommends the same videos to a group of users who share similar interest, thus guiding users’watching behaviors. Besides, it can reducing system load combining with P2P technology. Therefore, group recommendation system is an useful recommendation that considers not only user satisfaction but also system performance. Our paper uses Pre-push dataset and PPTV streaming VoD system logs. To reach the goal of group recommendation, we apply community unfolding technology. Meanwhile, to solve the problem that the results of recent methods are always ununiformed, we propose an improved community unfolding algorithm. Besides, we group recommend videos based on proper group recommendation strategy. More specific, works and contribution of the paper include:(1) Analyze the behavior of users and statistic features of users and videos in PPTV dataset. It shows that both user activity and video popularity subject to power law. Besides, we mine implicit user interest information from their behaviors, and calculate their Jaccard similarity.(2) Because of lacking video duration and user rating information, we design an implicit rating strategy based on statistic by mapping users’watching time to rating. Experiments show that it works very well in K Nearest Neighbors Algorithm.(3) We discuss the relationship between Peer amounts and download bandwidth as well as Peer amounts and P2P ratio in P2P system. The Pre-push dataset tells that P2P ration increases with Peer amounts at first, but step into a platform stage after some threshold.(4) We propose an improved community detecting algorithm. It can obtain more uniform results. Experiment shows that it can significantly improve uniformity by sacrificing little modularity.(5) We propose a group recommendation system framework combining user interests and system performance. It can balance both of them, effectively reducing system load.