Leveraging cellular network signaling data for origindestination matrix construction and travel patterns extraction in large-scale areas

Abstract

Human movements are the major mechanisms behind various spatial and temporal phenomena, such as urban commuting, transportation of goods, spread of influenza, and diffusion of automobile pollutants. Thus, the study of human travel behavior is crucial across a broad spectrum of research domains and for several applications, including urban planning, transport forecasting, epidemiology, and ecology. The study of human mobility requires a precise observation of individuals’ movements across territories, which is usually accompanied by major issues. For transportation planning and management, traditional data acquisition methods, such as household face-to-face or telephone-based interviews and roadside surveys, have become burdensome and inefficient. They only provide a snapshot of people’s mobility since they cover a limited sample of population and short observation periods. Also, they require considerable resources in terms of time and cost, and may not produce updated data. Relying on these traditional harvested travel data, transport demand models are facing many challenges as well to be easily applied. They require accurate individual trip information to correctly predict people’s mobility. At the same time, this requirement makes them difficult and expensive to deploy within multiple study areas. Without high-quality data, these models perform poorly and cannot be updated regularly. In addition, transport models are often calibrated for specific locations, and focus on specific environments. Because of the above, our understanding of the temporal and spatial distribution of trip flows at national or regional level is still very limited. The lack of deep knowledge about travel behaviors and their changing dynamics at large geographical scale may risk to misrepresent reality, underestimate effects, or simply lead to inefficient operations and decision-making biased by outdated baselines. With the recent advances in communication technologies, computation and storage capabilities, new data sources have emerged. Among the emerging sources, mobile phone network data are currently one of the most promising human mobility data sources. Alongside this data revolution there has been significant social change in how individuals plan movements, and hence in their travel behavior. This dissertation attempts to complement the conventional travel data collection methods and to leverage universal promising mobility data sources, capable to fundamentally shift the state-of-the art approaches of travel demand modeling. The first objective of this thesis is to design and develop a well-suited methodology to estimate the travel demand of a large-scale territory based solely on cellular network signaling footprints of mobile phone users moving within it. Fully comprehensive and elaborate data-driven approach is proposed to transform rich cellular signaling data into origin-destination (OD) flow matrix, valuable input for most transportation studies. In the resulting methodological framework, many important aspects are targeted: identification of residents, computation of novel cell phone activity indicators to filter relevant mobile users, trip extraction, and the introduction of an enhanced expansion method using zone-based scaling factors. As a major contribution of this work, we apply the proposed methodology in a real case study and present an extensive validation step which compares, at multiple levels, the obtained OD trip flows with an OD matrix estimated from a local travel survey. To this end, cellular signaling traces issued from 2G and 3G networks in the Rhone-Alpes region, France (about 44,000km2) have been leveraged. They involve more than 2 million subscribers and cover a period of 24 hours. The validation experiments show very promising results with high measured similarities (regression coefficient of about 0.95), opening up possibilities for the combination of rich signaling data with official travel surveys in unforeseen ways. The efficiency of our approach (i.e., relatively small signaling dataset required as input, cost and time effective method comparing to conventional surveys) makes it generally applicable and ready to be used in practice. The second objective is to explore the potential of mobile signaling data to provide useful knowledge about the travel dynamics and mobility patterns. Through a novel and unique pipeline that extends the framework previously developed, the dynamic travel demand is first estimated at large-scale level. The inferred hourly trip flows are then investigated to classify the emitted travel flows into distinct mobility profiles using an unsupervised learning algorithm. This spatial clustering reveals three meaningful temporal patterns of moving persons within heterogeneous environments. These three clusters represent 88% of the analyzed zones. The interpretation and evaluation of the estimated hourly travel demand as well as the generated mobility patterns are conducted using a real world 24-hour signaling dataset (the same used in the first experiments), land use and travel survey data. This is done after applying a novel correction solution on signaling data-based trips. The comparison analyses uncover interesting similarities (correlation coefficient between 90% and 95%) with the survey-based demand. This represents a strong added value for strategic planning and management of transportation networks since the dynamics of the daily travel patterns within a territory cannot be extracted from traditionally collected mobility data.