Recommender systems are an effective tool to resolve information overload by enabling the selection of the subsets of items from a universal set based on user preferences. The operation of most of recommender systems depends on the prediction ratings, which may introduce a degree of uncertainty into the process of recommendation. However, systems equipped with only two strategies lack the flexibility to address such uncertain decision-making problems. Thus, the presence of far-fetched recommendations accompanied by uncertainties often decreases recommendation quality. To resolve this issue, a three-way recommendation model based on a novel shadowed set is proposed in this paper to reduce decision-making risk and improve quality. To this end, a neighborhood rough set model is first introduced into three-way recommendation to determine similar user to active users with respect to the original rating decision system. This helps to avoid the uncertainty generated during the assignment of prediction rating. Subsequently, the optimal neighborhood radius is defined to overcome the subjectivity associated with the construction of the aforementioned neighborhood with a subjective parameter. Following this, a novel shadowed set model, based on neighborhood memberships of boundary users, is proposed to partition all users into positive region, negativity region and boundary region. This facilitates the adoption of different decisions by recommender systems for users in different regions. Finally, the effectiveness and reliability of the proposed model are verified on two Movielens datasets via comparison analyses.