Memory-based approaches for collaborative filtering predict users` preferences on items. Its prediction is based on average ratings given by other users who have similar rating patterns. The prediction accuracy of memory-based approaches is usually better in case of having sparse datasets. It is because the memory-based approaches can capture local associations in the datasets. However, the memory-based approaches have scalability problems in processing large-scale datasets because they use full datasets for every time when they make rating predictions. Especially, the similarity measurement between users consume most of the time during the recommendation process. The computation time of the memory-based approaches increase exponentially when the number of users is increased. In this work, we propose a filtering method to reduce the time for similarity measurement by considering only the users who showed their preferences on popular items. Furthermore, we consider users` preferences on the types of items (e.g. genres for movie items) rather than individual items to reduce the number of users to compare in the similarity measurement. By considering both popular items and type information, we could decrease the running time of the recommendation process by 70% in average while maintaining the accuracy of recommendation. Our approach outperforms existing memory-based approaches that use clustering methods. The evaluation was conducted with a real-world dataset from MovieLens.