Personalized Video Fragment Recommendation

In the mass market, users' attention span over video contents is agonizingly short (e.g., 15 seconds for music/entertainment videos, 6 minutes for lecture videos, etc.), from a video producer's or platform provider's point of view. Given the huge amounts of existing and new videos that are significantly longer than such attention spans, a formidable research challenge is to design and implement a system for recommending just the specific fragments within a long video to match the profiles of the users.In this paper, we propose to meet this challenge based on three major insights. First, we propose to apply Self-Attention Blocks in our deep-learning framework to capture the fragment-level contextual effect. Second, we design a Video-Level Representation Module to take video-level preference into consideration when generating recommendations. Third, we propose a simple yet effective loss function for the video fragment recommendation task. Extensive experiments are conducted to evaluate the effectiveness of the proposed method. Experiment results show that our proposed framework outperforms state-of-the-art approaches in both NDCG@K and Recall@K, demonstrating judicious exploitation of fragment-level contextual effect and video-level preference. Moreover, empirical experiments are also conducted to analyze the key components and parameters in the proposed framework.