Trip distribution modeling with Twitter data

Highlights

• Traditional and social media data are integrated to predict commuting trip distribution.

• The predictive performance of the traditional gravity model and ML techniques is evaluated.

• The potential of Twitter data in trip distribution modeling at fine spatial scale is explored.

• The most significant predictor variables are determined.

Abstract

Integrating both traditional and social media data, this study compares the performance of gravity, neural network, and random forest models of commuting trip distribution in New York City. Trip distribution modeling has primarily employed traditional data sources and classical methods such as the gravity. However, with the emergence of social media during the past decade, the potential for integrating traditional and social media data while utilizing new techniques has been identified. Our findings indicate that the random forest model outperforms the traditional gravity and neural network models. Population, distance, number of Twitter users, and employment were identified as the four most influential predictors of trip distibution by the random forest model. While Twitter flows did not enhance the models' performance, the importance of the number of Twitter users at work destinations implies the potential for using social media data in travel demand modeling to improve the predictive power and accuracy.

Introduction

Worldwide increases in traffic congestion and air pollution in urban areas presents a need to better understand mobility patterns of urban populations and their travel demands (Shirzadi Babakan, Alimohammadi, & Taleai, 2015; Yang, 2013). A number of studies have examined either individual or collective mobility patterns at different spatial scales (i.e., Beiró, Panisson, Tizzoni, & Cattuto, 2016; González, Hidalgo, & Barabási, 2008; Hawelka et al., 2014). Mobility information of individuals can be aggregated to study the frequency of travel between different regions, as represented by origin-destination (OD) matrices (Barbosa et al., 2018). An OD matrix provides population flow patterns (trip distribution) studied for diverse purposes such as traffic forecasting, resource allocation, prediction of migration flows, and epidemic spreading (Beiró et al., 2016; Pourebrahim, Sultana, Thill, & Mohanty, 2018). Therefore, improving flow estimations has become critical across various domains of application (Barbosa et al., 2018).

Various trip distribution models have been developed over past decades to estimate population flows with greater accuracy (de Dios Ortuzar & Willumsen, 2011; Roy & Thill, 2003; Simini, González, Maritan, & Barabási, 2012; Wilson, 1970; Wilson, 1998; Zipf, 1946). Most of these models are heavily dependent on conventional data such as population censuses or travel diary surveys. The emergence of social media and location-based services in recent years has introduced new opportunities to the field of transportation (Yang, Jin, Cheng, Zhang, & Ran, 2015). Geospatial big data such as taxi trajectories, mobile phone records, and social media messages have attracted scholars to observe, understand, and visualize (i.e., Karduni et al., 2017) human activities in cities at fine spatio-temporal scales (Liu et al., 2015). These data significantly improve the visualization of human mobility patterns, yet there is a need to better understand and contextualize them in different steps of the travel demand modeling framework (Anda, Erath, & Fourie, 2017).

Traditionally, the gravity model and its derivatives have been used as the most reliable approach to predict trip distribution at fine spatial scales, such as commuting flows within cities (Lenormand, Bassolas, & Ramasco, 2016). The potential for developing hybrid approaches that integrate the vast volume of social media data with the gravity model has recently been noted (Beiró et al., 2016). While traditional models have used statistical methods rooted in sound mathematical foundations, they have been unable to account for nonlinearities and other irregularities in data (Golshani, Shabanpour, Mahmoudifard, Derrible, & Mohammadian, 2018). To alleviate these and other issues, Machine Learning (ML) techniques have been applied in different urban and transportation domains (i.e., Ghasri, Hossein Rashidi, & Waller, 2017; Karimi, Sultana, Shirzadi Babakan, & Suthaharan, 2019). A significant body of literature exists at this time where various ML techniques have been evaluated on their ability to model travel demand, such as artificial neural networks (ANNs) (i.e., Ding, Wang, Wang, & Baumann, 2013; Mozolin, Thill, & Lynn Usery, 2000; Pourebrahim et al., 2018; Tillema, van Zuilekom, & van Maarseveen, 2006) and tree-based ensemble methods (i.e., Ghasri et al., 2017; Rasouli & Timmermans, 2014). While random forests (RFs) (Breiman, 2001) have been identified among the most advanced and most efficient ensemble methods for data classification and regression (Ghasri et al., 2017), they have so far been used in only a few studies of travel demand. Ghasri et al. (2017) and Rasouli and Timmermans (2014) have reported promising results with RF modeling of trip generation and modal split, yet their suitability and usefulness in trip distribution analysis remains to be thoroughly assessed.

Given the current state of research, our objective is to compare the performance of gravity, neural network, and random forest models of commuting trip distribution while combining both traditional and social media data. We also evaluate how information on personal mobility derived from social media affects commuting trip distribution by identifying the importance of different variables. To the best of our knowledge, this paper is one of the first to use machine learning approaches in trip distribution forecasting with social media data. The main contributions of this paper are threefold: (1) revealing the potential of social media data in trip distribution modeling at census tract level; (2) using machine learning techniques to predict trip distribution at census tract level; and (3) comparing the performance of gravity, neural network and random forest models to identify the best model for predicting trip distribution at census tract level. The paper is organized as follows. The review of related work is provided in section 2, followed by a presentation of the study area, data sources and methodology in section 3. Results are presented in section 4, with a discussion and concluding remarks in sections 5 and 6.