[4Yin2] Interactive session 2

[4Yin2-35] Supervised Embedding with Q-values for DQN Recommender Systems

〇Shota Inoue¹, Kazushi Okamoto¹ (1. The University of Electro-Communications)

Keywords:Recommender System, Reinforcement Learning, Deep Q-Network, Embedding Learning, Supervised Learning

Item/user embeddings are one of the important challenges in the deep reinforcement learning based recommender systems. Many studies leave embeddings as pre-trained and fixed. However, such fixed embeddings cannot model user preference shift, and pre-training of embeddings is difficult for less observed data. This study proposes a framework which updates embeddings by supervised signal with weighted Q-values, and validates recommendation accuracy, long-term rewards, and characteristics of embedded vectors via experiments by using real-world datasets. The results suggest that the proposed model without pre-training achieves the same accuracy level as other baselines and can obtain embeddings that users and items are widely distributed in the feature space.