A STUDY ON CONTENT-BASED VIDEO RECOMMENDATION

Video streaming services heavily depend on the video recommender system to help the users discover videos they would enjoy. Most existing recommender systems compute video relevance based on users implicit feedbacks, e.g., watch and search behaviors. For example, one can use Collaborative Filtering based methods to model the user-video preference, and compute the video-video relevance scores. However, when a new coming video is added to the library, the recommender system has to deal with the cold-start problem, i.e., to boot-strap the video relevance score with very few user behavior with respect to the newly added video. To solve this problem, in this paper we propose a content-based video recommendation approach by taking the advantage of deep convolutional neural networks to alleviate the cold-start problem. The proposed approach works well, especially in the case of serious data incompleteness. In addition to the vision feature, we also conduct extensive evaluation on video meta-data, and audio features.