Identifying Local Spatiotemporal Autocorrelation Patterns of Taxi Pick-ups and Dropoffs

Abstract

GIScience 2016 Short Paper Proceedings Identifying Local Spatiotemporal Autocorrelation Patterns of Taxi Pick-ups and Drop-offs Song Gao 1 , Rui Zhu 1 , Gengchen Mai 1 STKO Lab, Department of Geography, University of California, Santa Barbara, USA Contact Email: sgao@geog.ucsb.edu Abstract Analyzing spatiotemporal autocorrelation would be helpful to understand the underlying dynamic patterns in space and time simultaneously. In this work, we aim to extend the conventional spatial autocorrelation statistics to a more general framework considering both spatial and temporal dimensions. Specifically, we focus on the spatiotemporal version of Getis-Ord's G * . The proposed indicator STG * can quantify the local association of adjacent features in space and time. As a proof of concept, the proposed method is then applied in a large-scale GPS-enabled taxi dataset to identify local spatiotemporal autocorrelation patterns of taxi pick-ups and drop-offs in New York City. 1. Introduction Nowadays, large-scale spatiotemporal data (e.g., taxi trajectories, phone call records, social media posts) become available, which provide rich information to support research on human behaviors, transportation, urban landscape, and human-environment interactions. However, discovering patterns hidden in large-scale spatiotemporal datasets is challenging and thus attracts a lot of attention from the GIScience community (Hardisty and Klippel 2010; Demsar and Virrantaus 2010; Scholz and Lu 2014; Claramunt and Stewart 2015). Spatial autocorrelation statistics, like Moran's I and Getis-Ord's G are commonly designed for identifying spatial autocorrelation patterns (Fischer and Getis 2009). However, there is a gap in building corresponding measurements for spatiotemporal autocorrelations. For example, although human movements and activities may vary over time across different places, the observed activity hotspots and movement flow might exhibit a pattern of spatial dependence. Also, ignoring the temporal dimension would not be sufficient to discover underlying spatiotemporal dynamics. Therefore, our work aims to contribute to extend the conventional local spatial autocorrelation statistics to include both spatial and temporal dimensions. As a proof of concept, the proposed method is then applied in a large-scale GPS- enabled taxi dataset to identify local autocorrelation patterns of taxi pick-up points (PUPs) and drop-off points (DOPs) in New York City. 2. Methodology Spatial autocorrelation measurements can be divided into two categories: global and local indices. Classic global indices of spatial autocorrelation include Moran’s I, Geary’s C, and Getis-Ord’s General G, while local indices of spatial association (LISA) can be established by transforming the global indices into corresponding local measurements (Anselin 1995). Spatiotemporal autocorrelation concept refers to the relationship between some variable observed in each of space-time settings and the association with its neighbors. In a previous work, Gao (2015) proposed three global spatiotemporal autocorrelation indices but didn’t describe how to decompose them into local versions. As an initial trial, this work focuses on extending Getis-Ord's G * (Equation 1) (Getis and Ord 1992) by adding temporal indexes into