Long-term engagement-oriented short video recommendation via deep reinforcement learning

Abstract

Short video platforms require recommendation models that can adapt to rapidly changing user interests while optimizing engagement beyond immediate click feedback. Existing collaborative filtering and supervised deep learning methods are effective for static preference matching and short-term click-through-rate prediction, but they are less suited to sequential recommendation scenarios where each recommended item influences subsequent user behavior. To address this limitation, this paper proposes a deep reinforcement learning (DRL) framework for short video recommendation. The recommendation process is formulated as a Markov Decision Process (MDP), in which the user state is constructed from historical interactions and multimodal item features, the action corresponds to selecting candidate videos, and the reward integrates immediate feedback with delayed engagement signals. A Deep Q-Network (DQN)-based agent is trained in an offline simulation environment based on the KuaiRec short-video dataset and is compared with representative baselines, including popularity-based recommendation, matrix factorization, and supervised DNN models. The experimental results show that the DRL agent better balances click relevance, session duration, and recommendation diversity, particularly in long-term engagement-oriented metrics. Ablation results further confirm that combining immediate and delayed reward signals is important for learning a robust recommendation policy. This study provides a practical framework for developing intelligent short video recommendation systems that shift the optimization objective from short-term response maximization toward long-term user value.