Enhanced Short Video Understanding by Integrating User Behavior and Multimedia Content Information

摘要

The focus of ICME 2019 Grand Challenge is short video understanding and recommendation system based on user-video interaction data and multi-modal video features, including visual, text, and audio features. This paper provides the solution of our team hanhan to track 2 of this challenge. We cast this problem as a binary classification problem and addressed it by careful feature engineering and gradient boosted decision trees. To fully exploit the implicit feedback and multi-model content information, we created truncated SVD-based and neural-net-based embedding features for users, videos and authors. Furthermore, ensemble of a collection of models that take into consideration the cold-start and imbalanced nature of the recommendation task can further significantly improve upon the best single model. By using the proposed approach, our team was able to attain the 1st place in track 2 of the competition.