Leveraging Representation Learning for Urban Prediction Tasks

Keywords: urban prediction, spatial scale, data quality, deep learning, representation learning
Abstract Type: Paper Abstract

Abstract

This work explores the use of deep representation learning for urban region prediction tasks with respect to trade-offs between spatial scale and data quality and availability. Urban representation learning aims to extract key features of a region to a numerical encoding which can be used for a variety of downstream prediction tasks related to human residents and lifestyle. Examples of these tasks are COVID-19 forecasting, demographic and socioeconomic prediction, point-of-interest (POI) type classification, land usage, and more. Standardized census entities, such as blocks, tracts, or counties, are often used to define regions for analysis for urban predictions. Methods for defining census boundaries are based on population and features of the environment such as roads or bodies of water. This leads to variation in the size and shape of the region, causing differences in data volume and availability of high quality data across regions. Furthermore, census-based spatial regions can have high heterogeneity, therefore tasks may benefit by defining higher resolution boundaries. We employ a multi-modal model which learns low-dimensional regional encodings that can be used to evaluate similarities between regions and aid downstream tasks. With these representations we investigate characteristics of defined regions, such as uniform or diverse composition, size of the region, and data availability, and the mechanisms in which these impact the learned representations and downstream tasks such as demographic prediction.