Authors: Di Zhu*, Peking University, Fan Zhang, MIT Senseable City Lab, Yu Liu, Peking University
Topics: Geographic Information Science and Systems, Spatial Analysis & Modeling, Geographic Theory
Keywords: big geo-data, graph convolutional neural networks, place, spatial prediction, deep learning
Session Type: Paper
Start / End Time: 3:20 PM / 4:35 PM
Room: Director's Row I, Sheraton, Plaza Building, Lobby Level
Presentation File: No File Uploaded
Place is a fundamental concept in geographical analysis. Predicting the unknown properties of a place relies upon both the observed attributes of the place and the characteristics of its contextual places or neighborhoods. However, since place characteristics are unstructured and the metrics for places' contexts could be diverse, it is hard to incorporate both places' characteristics and places' connections in traditional spatial prediction models. To bridge the gap, we introduce graph convolutional neural networks (GCNNs) to model places as a graph, where each place is formalized as a node, places' connections are represented as the edges, and place characteristics are encoded as node features. A case study was designed in Beijing urban area to predict for the unobserved place characteristics based on the observed place characteristics and places' connections. A series of comparative experiments was conducted to highlight the influence of different place connection measures on the prediction accuracy, and to evaluate the predictabilities across different characteristic dimensions.
This research enlightens a promising future of GCNNs in formalizing places in geographic knowledge representation and reasoning.