Contextual bandits to increase user prediction accuracy in movie recommendation system

Cold-start problems are inevitable phenomena where recommendation systems fail to accurately predict users’ favour and cause the loss of new users. The typical Multi-Armed Bandit (MAB) models are widely adopted as recommendation systems to solve cold-start problems, but standard MAB takes much more recommendation trials than new user’s tolerance. This study adopts Contextual Multi-Armed Bandit (CMAB) to alleviate such situations and compares the performance of CMAB and typical MAB models at an early stage of the cold phase. Overall, CMAB generated better results in 15 trials in terms of cumulative regret and discounted cumulative gain. The optimal number of groups is 10, which alleviates cold-start problems efficiently, and sustains the efficiency of the off-line recommendation system under collaborative filtering. This paper suggests a possible selection of CMAB for recommendation systems to alleviate the cold start problem and estimates the tuned parameters for the MovieLens dataset. The evaluation metric in this paper provides a possible method of analyzing the general performance of a hybrid recommendation system, instead of adopting multiple evaluation metrics respectively, these metrics also provide estimates of the optimal value of parameters.