Interactive Recommendation Algorithm based on Combination of DFM and Offline Reinforcement Learning
Shangxian Fang, Chuanbao Ren, Wei Shen, Wei Xiong · 2023
In recent years, with the widespread use of short video applications such as TikTok, interactive recommendation systems have received wide attention. Unlike traditional recommendation systems, in interactive recommendation, the system not only needs to pay attention to the user's current interest, but also needs to capture the user's interest changes caused by the current recommendation, optimize the recommendation policy, increase the user's usage time, and obtain higher reward. Therefore, traditional deep learning recommendation methods cannot adapt to the interactive recommendation scenario, while offline reinforcement learning can use a large amount of user historical data, rely on the agent and the environment to continuously interact and learn autonomously, and obtain a personalized recommendation policy that satisfies the user. This paper proposes an offline reinforcement learning algorithm called RORL, which uses DFM to process the high-dimensional sparse features of user historical data, and then uses the BCQ algorithm to train the optimal recommendation policy. Compared with the four baseline algorithms, our model performs better than them on multiple metrics, such as Precision@k and MAP, and further ablation experiments show that offline environment and factorization machine can effectively improve the recommendation effect.