Using a Human Mobility Pattern Prediction Model to Estimate Trip Distribution Matrix

Travel demand forecasting is an important topic in transportation planning. Understanding and modeling the travel demand has numerous applications in designing urban infrastructures, managing the spread of diseases, monitoring the dispersion of computer viruses, urban planning and policies, spatiotemporal analyses in GIS, and Location-Based Services (LBS). Traditionally, in order to predict the travel demand, a four step model is used, of which the second step is called trip distribution. The output of trip distribution step in this model is termed Origin-Destination (OD) matrix. The elements of this matrix indicate the amount of trips departing from origin zones to destination zones. The OD matrix is considered as an important input in various spatial analyses in Geospatial Information Systems (GIS). The most essential part of trip distribution is the model used for OD matrix estimation. Up to now, various models such as gravity have been introduced to estimate the trip distribution. Recently, some parameterized and non-parametric models of human mobility pattern prediction, also known as spatial interaction (SI) models, have been developed. Among them are the rank-based (parameterized), radiation (non-parametric), and PWO (non-parametric) models. These models can be applied to a broad range of scales, from within a house or stadium, to a city, country, or even the whole earth. The probabilistic form of these models is the same as OD estimation models. In addition, in these models, computational mechanisms of trip distribution are not limited and different behavioral and decision-making characteristics of people are also taken into account. In this paper, the applicability of PWO, radiation and rank-based models in OD matrix estimation is addressed. As a case study, the rank-based model has been applied for Manhattan, New York City (NYC) and the results have been evaluated. Manhattan is one of the most important trade centers in the world and its mobility rate is remarkable. In order to calibrate the models in which adjustable parameters are appeared, the Hyman method was employed. Hyman method is a repetitive algorithm which uses secant procedure to minimize the difference between the real trips’ average distance and the modelled trips’ average distance. Also, in order to balance the resulting matrix, another process is needed. In this paper, the Furness method has been used. For the purpose of evaluating the results, trajectory data of taxi vehicles within NYC was employed. These dataset are published monthly by NYC Taxi and Limousine Commission (TLC). To capture a more complete pattern, the trajectories of yellow- and white-colored taxis were combined together. Finally, we used Sorensen Similarity Index (SSI), Regression Analysis, and a visual measure, termed sparsity pattern, to examine the model. The results show that the rank-based model can predict trip distribution up to 67 percent according to Sorensen Similarity Index. Additionally, the r-square measure obtained from regression analysis is 0.32 that shows a good agreement between estimated and ground truth matrices. Taking the huge volume of point data being regressed into account, this value shows a good agreement between modeling results and real trips. These results show the potential of recently introduced spatial interaction models in trip distribution estimation.