Abstract: Aiming at the problems of the limitations of  data sparsity and “cold start” on collaborative filtering and the inapplicability of the existing collaborative multi-armed Bandit algorithm to nonlinear reward functions, we proposed a neural Bandit recommendation algorithm COEENet, which combined collaborative filtering. Firstly, it adopted a dual neural network structure to learn expected  rewards and  potential gains. Secondly, we considered the collaborative effect of neighbors. Finally, a decision-maker was constructed to make the final decision. The experimental results show that the proposed method is superior to the four baseline algorithms in cumulative regret, and has a good recommendation effect.

Key words: collaborative filtering, multi-armed Bandit algorithm, recommendation system, cold start