Truncated SVD-based Feature Engineering for Short Video Understanding and Recommendation

Short video app, like TikTok, has received wide acclaim due to the prevalence of social media and the availability of recording devices such as mobile phones. Moreover, with the advent of the big data age, the use of historical user behaviors from multi-modal resources plays a pivotal role in the video recommendation system. In the ICME 2019 Short Video Understanding Challenge, participants are asked to predict whether a user will finish and like a specific short video along with its multi-modal features, i.e., the problem is formulated as a click-through rate prediction task. In this paper, we present an ensemble of unconventional models to the task, including tailored neural networks structure based on Compressed Interaction Network (CIN) and Gradient Boosting Decision Trees (GDBTs) using classic SVD-based features. We achieved a weighted AUC score of 0.8029 and 0.8037 on the Public and Private Leaderboard of track2, respectively, and ended up with the 3$^{rd}$ place in the competition.

Download paper here

Recommended citation: Tsun-Hsien Tang, Kuan-Ta Chen and Hsin-Hsi Chen (2019). “Truncated SVD-based Feature Engineering for Short Video Understanding and Recommendation.” In Proceedings of IEEE International Conference on Multimedia and Expo 2019, Short Video Understanding Challenge, July 8-12, 2019, Shanghai, China.