Automatic latent street type discovery from web open data

Abstract

Street categorization is an important topic in urban planning and in various applications such as routing and environment monitoring. Typically streets are classified as commercial, residential, and industrial. However, such broad categorization is insufficient to capture the rich properties a street may possess, and often cannot be used for specific applications. Previous works have proposed several advanced street categorization systems. However, most of these systems rely on manual analysis and design, which requires significant effort. In this paper, we propose a method for automatically discovering latent street types from multi-modal Web open data. We utilize data of different modalities including microblog tweets, Foursquare venues, and Google Street View images. The model we propose considers both coherence within each modality and association between modalities. Based on the San Francisco city data, our quantitative evaluation shows superiority of the proposed method in terms of coherence and association. In qualitative analysis, we show that the street types discovered by our method correspond to the official street plan. We also show an example application in which the discovered street types are used in crime prediction.