| Ever since the entrance of Web 2.0 era, the number of websites with characteristic of UGC(User Generated Content) has been increasing constantly, among which service of online video stands out. Unlike traditional video websites which based on movies and tv series like Netflix and Hulu etc., video websites with characteristic of Web2.0 like You Tube and Youku only provide hosting service with all of the videos generated and uploaded by the users themselves. This approach expands the content of video websites greatly. Meanwhile, users have all kinds of options, which results in costing more time and energy to find their real interests. One of the solutions is the personalization recommendation system, which can build user’s interest model based on user’s log of history behavior and extract information which user may get interested in automatically and push them onto the user on its own initiative. In that way, not only is the time of user’s obtaining information cut down dramatically and user’s experience promoted, but also websites’ income and scale of influence get a lot of help in advancing forward. As a result, the personalized recommender system has got bigger and bigger scale of application in famous websites such as Amazon, Netflix and Douban.There are still a lot of problems in the scene of the video recommender system because of allowing users to generate and upload their own videos. First, online videos are of complex categories and abundant dimensionalities. As a result, we can’t classify internet videos based on some explicit standards to generate video recommendations. Second, online videos vary a lot due to the irregularity of users’ spontaneous activities. Many online videos are lack of descriptive meta data, even with the existence of disorderly and unsystematic meta data. Last but not least, subject to the multi-media property of online videos, there are not existing reasonable and effective approaches of machine learning to extract content features directly from online videos themselves, which brings a lot of challenge to the similarity measurement of online videos and user interest model building.Considering the particularity of online video, we proposed a new recommender method, VRFCL(Video Recommender Fusing Comment Analysis and Latent Factor Model). Based on the user’s video comment, VRFCL first work out user’s sentiment tendency towards one particular video using text sentiment analysis technology and then build the virtual rating matrix on the basis of the sentiment tendency score calculated before, which makes up for the sparsity of explicit rating data. On the other hand, the comment texts always contain some objective analysis of video comment users contribute, which is very valuable. With help of technology of document analysis, we can extract key words of comment text as tags, which not only can make up for the lack of metadata of online video, but also has a lot of help in user interest building. At the same time, because of online video’s characteristic of high dimensionality and content diversity, latent factor is introduced in VRFCL to extend the traditional binary User-Term relation to triple User-Factor-Item relation based on the virtual rating matrix, which can promote the novelty and coverage of recommendation result produced by recommender system. In order to verify the effectiveness of VRFCL, we did a lot of experiments based on comment text set from You Tube, and evaluated the performance of the different recommender algorithms from the perspective of the different evaluation standard. The experiment results showed that VRFCL not only performed good on sparse rating data set, but also achieved improvement of 10%-20% degree in recommender precision compared to other recommender algorithms. |
PDF Full Text Request