In considering of the dynamic variations of user’s preference and item’s popularity, sequential recommender system (RS) has attracted much attention in recent years. In general, the sequential interactions between users and items will lead to both multilevel recommendation information (RI) in the space dimension and multi-step recommendation in the time dimension. To better capture the dynamic variations of user’s preference and reduce the recommendation cost, this paper proposes a novel sequential recommendation strategy from the temporal-spatial perspective. Firstly, in view of the temporality of user-item interactions, we design a granulation method based on recurrent neural network (RNN) to construct the multilevel RI. Then, in the light of the temporality of user’s preference and item’s popularity, we present a temporal-spatial three-way recommendation strategy (TS3WR) to realize the multi-step recommendation. Finally, by integrating the time factor with space factor, a temporal-spatial three-way recommendation based on recurrent neural network (RNN-TS3WR) is proposed to realize the recommendation with lower decision cost. Extensive experiments on three Movielens datasets verify the feasibility and effectiveness of our proposed methods, and demonstrate the advantage of our recommendation strategy in both recommendation cost and recommendation quality.