Practice of a New Model Fusion Structure in Short Video Recommendation

Abstract

In recent years, the popular short video content understanding and recommendation technology has become a research hotspot. This paper presents a new mixed short video recommendation algorithm based on the content, the use of FM, FMM, DeepFM DeepDCN model) and the short video content (visual, text, user interaction), and other characteristics of the training, at the same time, using a combination of two hybrid model of strategy for learning, for users to return to a short video recommend high quality results. In order to evaluate the proposed algorithm, experiments were carried out on the short video content understanding and recommendation competition data set on Biendata open competition platform, which was derived from the TikTok (douyin overseas edition) short video APP owned by bytedance company, which included two tracks: yayi and yayi. Participants were asked to model a user's interest through a video and user interaction data set, and then predict the user's click behavior on another video data set. Based on the analysis of root race results, the algorithm proposed in this paper effectively improves the performance of short video recommendation algorithm.