Learning Walkability Representation via Hyperbolic Embedding of Road Networks

Abstract

Walkability is the measure of an area's friendliness by walking to the nearby amenities, such as schools, shops, restaurants and entertainment facilities. It is an a ecting factor for prediction of the Hedonic pricing regression model in real estate market. Walkability has been increasingly gaining attention in the data science eld as it provides social, environmental and economic bene ts. However, the most popular walkability measurement, Walk Score, is a distance-based technique which su ers from computational expensive and time-consuming issue due to the complexity of large road networks in modern cities. Moreover, modern cities suggests a graph with the core-periphery structure which often imply hierarchical information. Some well-developed areas form the cores of a city containing massive amenities, while periphery regions are less accessible to amenities [60, 3, 2, 49]. With such inspiration, we treat amenities as the top-level nodes of hierarchies and model the walkability of properties on hierarchical structures. Nonetheless, it is impractical to directly implement the above formulation as there are tens of thousands of amenities in an urban city. Hence, we further utilized the representation learning technique to learn a latent vector representation of each residential property to preserve its proximity to all amenities. Current representation learning methods learn embeddings in Euclidean vector spaces, which do not account for latent hierarchical structure. Therefore, we present HyperWalk for measuring walkability by embedding the large road network into Poincar e ball hyperbolic space. Our technique utilizes the hyperbolic geometry to learn a representation that captures hierarchy of large road network. This approach embeds road segments with high walkability near the center of the Poincar e ball, while less walkable road segments are pushed toward the edge. We successfully demonstrated that our technique is able to capture the latent hierarchical relationship between amenities and residential properties and learn walkability that improve the prediction performance of the Hedonic pricing regression model.