Online DRQN-Based E-Commerce Recommendation

Abstract

E-commerce recommendation systems, which have the advantage of dynamic interaction over traditional recommendation algorithms, use reinforcement learning for virtual human-e-commerce interaction to model the interests of users. In this paper, we first embed users and items, and then pass them through an LSTM network. Unlike DRQNs, we do not require image recognition, so we do not use convolutional neural networks and instead only consider training these RL-based recommender systems via LSTM in a RecoGym environment ^[1]. In this paper, we use RecoGym to generate artificial data, rather than real data, in consideration of the requirement to protect user privacy and to be more easily compatible with a reinforcement learning environment. A combination of Long Short Term Memory (LSTM) and DQN is deployed. This compensates for the fact that DQNs handle longer sequences of actions or achieve "relaxed Markovian" learning across sequences. Similar modelling effects were demonstrated in ^[2][3][4], but this paper uses a combination of reinforcement learning algorithms and recurrent neural networks and achieves better error convergence rates and recommends items to users more efficiently and accurately.