We investigate a novel collaborative filtering bandit algorithm using neighborhood-based aggregation. Our method determines pairwise user similarities from past interactions and uses them to personalize the predictions of an adapted (contextual) bandit algorithm. Personalization is achieved by reweighting past observations, thereby skewing the aggregated information towards interactions with similar users. Only taking into account similar users is not enough however and we explicitly incorporate global information to enable good performance in cold-start situations. Experiments on three real-world datasets and across different scenarios show that our algorithm performs competitive against state-of-the-art methods while being conceptually more intuitive.