Historical Data Embedding Using Hopfield Networks in Reinforcement Learning-Based Recommender Systems

다국어 초록 (Multilingual Abstract)

In this study, we investigate a novel approach to utilize all past information in a reinforcement learning-based recommender system instead of using only limited current information. Specifically, we aim to enhance the performance of recommender systems by using FrozenHopfield (FH), an advanced model of Modern Hopfield Networks, which can consider past interactions between users and systems as states in a reinforcement learning model. To this end, we propose the TH-Rec model, which integrates FH with Transformer-XL that can effectively compress past information, examine their applications in recommender systems, and study how to apply them to a reinforcement learning-based recommender system model. In addition, we validate our approach using the KuaiRec dataset from Kuaishou, a popular short video platform, and compare the TH-Rec model's performance against other existing recommendation algorithms. The experimental outcomes indicate that the TH-Rec model significantly surpasses the baseline models, highlighting the advantages of using Hopfield Network in reinforcement learning-based recommender systems for historical data embedding. This research extends the potential of Hopfield Networks in the field of recommendation systems and reinforcement learning and is expected to make important contributions to related research in the future.