Machine learning methods for Urban Flows: spatial effects in Origin-Destinations

The movement of the people, goods and information has always been a major interest of the social scientist. This research area plays a pivotal role in transport planning, urban planning, government policy and a multitude of other agendas. With the urban population increasing every year, the development of more sophisticated methods of how to track and predict these movements have become a necessity. Predicting human flow has important benefits for health services, transportation or migration, but poses many theoretical and practical issues. In the 1980s, Fotheringham (1983; 1986) claimed that the nature of human flows is primarily influenced by individual decision-making process, which can be unique for each person according to many factors that influence them in their everyday life. Advances in the availability of open data and computing power, are facilitating the ability for more sophisticated and complex analytical methods to predict this research agenda.
Following Newton's laws of gravity, Wilson (1970) described a base framework for spatial interaction modelling (SIM) and entropy maximization. Since then, many authors have contributed on innovative designs of SIMs and this development has played a significant contribution to spatial analysis in the social science literature (Fischer, 2009). Fotheringham (1983) claims that traditional SIMs are often misspecified because they explicitly ignore occurrence of spatial structure effects. Therefore, he introduced competing destination models that reflect the hierarchical structure of human decision-making, incorporating entropy maximization to achieve more correctly specified model. More recently, spatial econometricians, such as LeSage (2008; 2009) and Anselin (1988; 2004), have developed a rich and comprehensive mathematical framework for modelling of spatial effects and interaction. These include introducing Bayesian spatial statistical models (Gelfand, 2012), SIMs (LeSage & Llano, 2013) and spatio-temporal models (Khana, et al., 2018). Griffith, et al. (2017) demonstrated the significance of LeSage and Anselin's mathematical framework for modelling and representing effects in SIMs using a Bayesian model specification. Griffith et al. incorporated spatial effects and found that the amount of predicted flows was six times higher than observed flows. This demands scientific attention, as the question of the correct way to account for "space" in SIM remains a large open question.
With the Big data and AI revolution, many have attempted to use machine learning (ML) to provide better flow prediction. This is very common in transportation geography and analysis. Ezzatabadipour, et al. (2017) build an algorithm for predicting flow patterns using deep learning methods and compared its performance with the traditional methods. In addition to Ezzatabadipour, et al. (2017), Polson & Sokolov (2017) and Yi, et al. (2017) have found strongly positive results in their confidence in value and accuracy of ML's accuracy and prospects for flow prediction. This has basis in prior work as well, Fischer (2001; 2009; 2010) described a comprehensive methodology for neural SIM. Together, geospatial artificial intelligence is an emerging field, grounded in a large amount of prior work, that stands to apply new technologies to more realistically model humal flows (VoPham, et al., 2018). The aim of this project is to apply new AI technologies such as neural networks or deep neural networks on population flow data, also defined as origin-destination data, to examine the spatial effects within the spatial interaction methodology and provide a flexible framework for predicting spatial flows. Using Convolutional Neural Networks (CNN) will be convenient method that can capture the dynamicity of the flow data in space but possibly also in time.