| Summary: | 碩士 === 國立臺灣科技大學 === 資訊工程系 === 96 === Content-targeted advertising is a popular advertising strategy. The goal of content-targeted advertising is to associate ads with appropriate web contents that can reach a large number of targeted customers. However, searching hot videos by analyzing video contents will cause higher False Positive Rate, due to the characteristics of videos: massive amounts, fast update, and redundancy. Besides, searching hot videos by analyzing insufficient time-series data causes lower accuracy, due to online video’s fast burst and obsolescence nature. For improving the accuracy of prediction, we utilize user social context to alleviate the variation of video content and to improve the insufficient data problem in early prediction stage. In this paper, the UISN is constructed to represent the hot videos’ tendency by modeling user interest relation. The main idea of the proposed system is to identify cohesive subgroups of users with similar interests, so that it can be utilized to predict possible online videos that most people might feel interested. Finally, the UISN is adapted to new change of user interest over time. By using UISN to enhance insufficient information in early prediction stage, the proposed system can effetely predict hot videos. In addition, UISN can alleviate hot video prediction inaccuracy caused by the characteristics of online videos. Furthermore, by adapting user interest change and filtering noisy users, the FP-rate can be controlled under 2%, in the meanwhile; video prediction accuracy is slightly decrease.
|
|---|
