FlowTime: Towards Continuous Generative Watch Time Prediction via Flow-based Personalized Priors
Researchers have introduced FlowTime, a novel method for predicting user watch time in short-video recommendation systems. This approach utilizes a one-step generative variational autoencoder to model complex, multimodal user interaction patterns more effectively than existing regression techniques. FlowTime aims to improve user engagement by offering a more accurate and efficient prediction of how long users will watch videos, outperforming current state-of-the-art methods in extensive testing. AI
IMPACT Enhances recommender systems by enabling more accurate prediction of user engagement with video content.