Transferable Contextual Bandits with Prior Observations

Cross-domain recommendations have long been studied in traditional recommender systems, especially to solve the cold-start problem. Although recent approaches to dynamic personalized recommendation have leveraged the power of contextual bandits to benefit from the exploitation-exploration paradigm, very few works have been conducted on cross-domain recommendation in this setting. We propose a novel approach to solve the cold-start problem under the contextual bandit setting through the cross-domain approach. Our developed algorithm, T-LinUCB, takes advantage of prior recommendation observations from multiple domains to initialize the new arms' parameters so as to circumvent the lack of data arising from the cold-start problem. Our bandits therefore possess knowledge upon starting which yields better recommendation and faster convergence. We provide both a regret analysis and an experimental evaluation. Our approach outperforms the baseline, LinUCB, and experiment results demonstrate the benefits of our model.