Abstract: Data mining on check-in data inlocation based social networks (LBSNs) is an important research direction of urban computing and smart city, and a critical task is to infer location semantic. The study of location semantics has attracted increasing attention in diverse fields due to its wide applications such as location retrieval, location recommendation, data preprocessing and so on. Established inference approaches tend to manually discover the spatiotemporal pattern of unique location as features for training classifiers. However, extracting valuable spatiotemporal patterns or features is a non-trivial task. In this paper, we propose a novel location semantic inference with graph convolutional networks (SI-GCN). We introduce node2vec and variational autoencoder to learn spatial and temporal features of location, respectively. Furthermore, we leverage graph convolutional networks to capture high order relations in user’s check-in activity with building a user-location bipartite network. And leveraging self-attention mechanism is allowed to distinguish contributions of the different nodes. Extensive experiments on several real-world check-in data sets show that our proposed framework outperforms than three state-of-art algorithms.