Learning urban region representations with POIs and hierarchical graph infomax

We present the hierarchical graph infomax (HGI) approach for learning urban region representations (vector embeddings) with points-of-interest (POIs) in a fully unsupervised manner, which can be used in various downstream tasks. Specifically, HGI comprises several key steps: (1) training category embeddings as the initial features of POIs; (2) interconnecting POIs with a graph structure and performing graph convolution to capture the uniqueness of each POI based on its spatial context; (3) aggregating POIs to the regional level using multi-head attention mechanisms, to consider the multi-faceted influence from POIs to regions; (4) performing graph convolution at the regional level to generate region representations, to incorporate the similarities between adjacent regions; (5) aggregating region representations to produce an embedding at the city level. The model is finally trained through maximizing the mutual information among the POI – region – city hierarchy, which facilitates the information from local (POIs) and global (city) scales flowing to the learned region representations, making them both locally and globally relevant. We perform extensive experiments on three downstream tasks, i.e., estimating urban functional distributions, population density, and housing price, in the study areas of Xiamen Island and Shenzhen, China. The results indicate that HGI considerably outperforms several competitive baselines in all three tasks, which proves that HGI could produce meaningful and effective region representations. In addition, the learned region representations based on POIs can potentially be used for reinforcing data representations from other modalities, e.g., remote sensing data. The implementation of HGI can be found at https://github.com/RightBank/HGI.