As an important branch of recommendation systems, POI recommendation recommends the next Point of Interest (PoI) to users, which can effectively solve the problem of data redundancy in the information era. Researchers are committed to mining spatial-temporal information to model users' dynamic preferences and behavior patterns. However, suffering from the inherent sparsity issue of check-in records, existing methods only focus on spatial-temporal information and ignore the semantic information hidden in the users' check-in sequences. This kind of semantic information often embodies close relationship of user-POI and POI-POI. To this end, we introduce POI category information and propose a category-aware network with long- and short-term preferences (LSCAN for short) for next POI recommendation. By modeling the competition/cooperation relationship between POIs, we can have a deeper insight into user behavior patterns. Furthermore, the dynamic changes of users' interests, i.e., long-term and short-term preferences, play a significant role when users make decision. Therefore, we design a dynamic fusion mechanism to combine them. Finally, we study the performance of the proposed method on three real-world datasets. The experimental results show that LSCAN consistently outperforms the state-of-the-art approaches for next POI recommendation.