The Research And Implementation Of Hybrid Personalized Recommender For VoD System

The explosive growth of video-on-demand service has led to the revolution of the waypeople obtain information all over the world in the past decade. The online delivery ofvideo content have been surging to an unprecedented level, along with the problem ofinformation overload. A promising way to address this problem is the personalizedrecommender technique. It is not only widely applied in industry, but also remains highinterest by academic.Unfortunately, by comparing and analyzing previous personalized recommenders,several limitation has been found both for collaborative filtering(CF) method andcontent-based(CB) method. Addtionally, certain performance highly related to userexperience, such as high real-time and diversity of recommendation, were rarely mentionedby academic. Inspired by the observations above, we proposed a framework seperatedvideo network construction from personalized recommendation. It firstly constructs a unifyvideo network based upon historical view data, original textual metadata and high-levellatent semantic using Hadoop Map-Ruduce. The weight of edges in this graph, that isrelevancy between videos, are the linear combination of relevancy based on item-based CF,video content and intelligent topic; To generate recommend indexes for each user withdiffirent preference, we then developed a server with real-time performance and linearscalability. It searches candidates in the video network based on users’ latest interest(i.e.seeds), determining the relevancy between candidates and seeds. At last, it limited thenumber of output items derived from same seed in order to broaden the topics convered byrecommendation.The experiments on Netflix dataset have indicated that our proposed framework has apretty high recall on average,9.92%, which significantly better than those of other models.It is also a quite qualified recommender in terms of average precison,26.33%, which ispreceded only by pure CF model. Generally, our famework has the highest Matthewscorrelation coefficient,13.88%, which represents the overall ability of a recommender. Inaddition, indicated by performance test, our real-time server has an extremely short respondtime and linear scalability. This application has already deployed on our socialized VoDsystem, Cloudmedia.