A Knowledge Graph-based Interactive Recommender System Using Reinforcement Learning

To achieve the personalisation recommendation, modem recommendation models should consider the user’s preference for item attributes and dynamic changes in preferences. 1RS has attracted attention to dealing with dynamic user preferences. However, current 1RS models share a common issue of sparse user-interaction data for training an effective recommendation policy. In this paper, we propose a knowledge graph-based interactive recommender system (KGIRS) to improve the recommendation by considering the users’ dynamic preference for item attributes’ weight. This interactive recommender system incorporates the knowledge graph as the source of the auxiliary information to increase the user-item interaction data efficiency and utilises the Q-learning technique to detect the user’s preference drifting. In this model, Q-learning is used to learn the user preference on item attribute dimensions and detect the user’s preference drifting through the user’s every feedback. A user’s every interaction with items is modelled as a ripple in the knowledge graph. Their previous interest and dimensional preferences activate their potential interest in items. It propagates along the links in the knowledge graph from the ripple centre with the learned speed. Then obtain the candidate recommendation items in the unit propagation time according to the user’s preference. Through the experiments on the real-world dataset MovieLens-1M, we demonstrate that KGIRS can achieve efficient, personalised recommendations.