Large Scale Data Management In Video Social Network

With the development of the Internet, the explosive growth of the information online and the increasing bandwidth available for each single user, videos have been growing to the largest media content behind everyone's browsers. In the same time, Web 2.0 has become the most prevailing concept on the Internet. And video social network is currently the biggest platform for video publishing, sharing and communicating. Since the resources are augmenting every minute, how to effectively manage them becomes one of the most interesting research problems nowadays.Managing the video content online can be summarized into dealing with the problems among three conceptual entities:the video hosting service provider, the user and the 3rd party supervisor. The service provider would seek for effective data organization methods in order to help itself to maintain the data and to assist the users to find their favorite content. The user-centered Web 2.0 concept requires the service provider to give personalized service. And the video recommendation based on users' interest is a must-have feature for successful web video sharing sites. Because the users cannot be restricted at all time, the 3rd party supervisor would request the service provider to check whether the user-generated content violates any copyright laws, and delete them if any. In all these problems, new challenges and opportunities are brought by the new video social network environment at the same time.Focusing on the problems discussed above, this dissertation mainly studies how these problems can be effectively solved given the relationship among videos, users and the information within the videos. The contribution of this dissertation lies in three aspects:1. We propose a video topic discovery algorithm based on the links between videos and the text information around the videos. First of all, we enhance the poor-quality text information available on the video sharing websites via related video links. Secondly, we iteratively use a graph-cut based topic extraction method to detect video topics, and use the result to adjust the graph generation process. At last, video-video response relationship is utilized in refine the topic discovery result. Based on the experiment, our algorithm performs better in the accuracy than some earlier algorithms and it can save much time compared to topic model when the number of topics becomes large. 2. We propose a personalized video recommendation framework based on emotion analysis and social constraint between users. Firstly, we design a hierarchical strategy using global video features to perform fast video de-duplication. Secondly, we quantize the user comments into user rating scores based on a dictionary-based emotion analyzer with text context. Thirdly, we apply social constraint on the last recommendation process to determine which specific video entry should be recommended to a specific user. Based on the experiment conducted on our dataset, our framework can offer an obvious performance increment.3. We propose a new research problem of parody videos, with a model to describe this kind of video, and a system using this model to do parody retrieval. First of all, we define parody videos, and introduce the problem of parody retrieval. Secondly, we propose a video volume description model aiming at the key characteristics of parodies. In this model, a video volume is described with bags of motion words ordered in their temporal positions. We also design a video matching algorithm particularly for calculating similarity between videos described with this model. At last, we also use the context information to achieve a better performance in our system. Our dataset contains 7 topics of parodies covering different situations. Based on the experiment, our video volume description model is effective, and the parody retrieval problem can be solved efficiently with its help.In the end of the dissertation, we look ahead to the future of the problem of data management in video social network based on concluding the contribution of our work.