Enhance Hyper-Actor Critic Framework in Session-Based Recommendation System With Graph Encoder

Reinforcement learning (RL) solutions such as the Hyper-Actor Critic model have effectively improved sessionbased recommendation systems because of their ability to capture and model the stochastic interaction between users and items. However, the sequential encoder in the Hyper-Actor is only able to exploit the temporal relationship from the history data for input state representation, which sometimes fails to capture the complex relationships between items in a specific session. To address this, we propose to extend the state representation with an additional Graph Encoder. The central part of our Graph Encoder is a global correlation graph which is based on a Graph Convolution Network and is pretrained in an unsupervised manner. The output graph embedding is then combined with the temporal embedding to utilize item-item relationships better and provide a more detailed representation. Experimental results on a real-world dataset highlight that our proposed module improves the recommendation system performance in both accuracy and stability metrics in comparison with other RL-based methods.