Mobility and spatial networks 1

While they are defined at the micro-scale of nodes and edges, a key feature of networks is their ability to capture structure across scales by modelling indirect interactions as paths and walks. This assumes that flow along edges is transitive and can be represented as a first-order Markov model. However, trajectory or path data from a variety of systems such as passenger itineraries, career paths, and clickstreams have higher-order dependencies that first-order networks cannot capture. In this work, we present a pipeline to create 'concise' networks that interpolate between first and second-order models by capitalising on patterns in second-order effects. At the core of our method, we use Convex Non-negative Matrix Factorisation to create state nodes that model latent modes of behaviour. Our method includes an interpretable measure to guide the complexity-quality trade-off, a prior to regularise the data, and is designed to process nodes in parallel. In addition to showing that it recovers planted modes of behaviour in synthetic data, we explore the insights created by our method into two real-world systems. First, from flight itineraries in the US, we uncover intuitive transit behaviour through national hubs, revealing their role in routing passengers across geographic regions. Second, we show that despite being much less complex, our concise network model captures the modular structure of a full second-order model of information flow on a social network.

Transportation infrastructure systems are the backbone of our society, fostering economic growth and driving innovation by conveying people and goods from origin to destination in a timely manner. Successful operation of these infrastructures depends on efficient routing of origin-destination demands given a fixed amount of resources available. Here, we extend the previous framework of the shortest-path percolation, which is a stylized model defined on a simple graph describing how the macroscopic structure of the transportation system collapses as resources are consumed to fulfill the origin-destination demands, considering a generalized structure of the transportation network with time-stamp, directionality, weight, and capacity: the minimum-cost percolation (MCP) model. We applied the MCP framework to the US air transportation system using publicly available data and conducted extensive experiments considering various costs and different origin-destination demands. Our results indicate that cooperation among airline carriers through shared flights can significantly improve the overall efficiency of the US air transportation system. These findings demonstrate the potential of the MCP framework to guide cooperative policies that enhance resource utilization and overall efficiency.

SMS-based alerts during emergencies are a highly effective communication tool due to their broad reach, low cost, and immediacy [1]. However, overly broad alerts can lead to over-evacuation, straining infrastructure, and eroding public trust. We analyzed the SMS-based alerts issued during Chile’s most devastating wildfire in 30 years, which struck Valpara\'so on February 2, 2024, causing 137 fatalities and affecting over 16,000 people. Using spatio-temporal data from a national mobile network and census-based proxies for socioeconomic status (SES), we categorized telecommunication towers into lower, middle, and higher SES groups. Official alerts from SENAPRED provided the timing and location of evacuation messages, while to analyse people's mobility, we employed XDR spatio-temporal data from a national mobile communication company.Our findings show that the first SMS alert prompted significant evacuation behavior across all SES groups, regardless of proximity to the wildfire (Figure 1a). Higher SES areas exhibited fewer connections (Figure 1b), suggesting greater capacity for proactive evacuation. Interrupted time-series analysis confirmed that the initial alert was the primary driver of evacuation, with higher SES evacuations influenced more by prudence than direct alert targeting (Figure 1c). Our findings show several interesting patterns; initial alerts elicit stronger responses, repeated alerts risk desensitization, and SES shapes evacuation patterns, with higher SES groups better positioned to act. These insights inform strategies to enhance emergency communication, ensuring precision and improved safety for all communities.

Contact networks, which describe individuals' physical interactions, have broad applications for understanding diffusion processes in social networks, from the spread of information to the transmission of infectious disease. Commonly, contact networks are segmented by individuals' demographic characteristics, showing the contact probabilities (or "mixing") between different demographic sub-populations. While demographic-stratification of contact networks can reveal important structure in the patterns of interaction within a population, demographic features do not fully capture behavioral differences between individuals. In this project, we build on existing research identifying latent behavioral patterns in individual activity, which looks below the level of individual demographics to identify common patterns of activity across individuals' lifestyles. Using detailed GPS mobility data describing individuals' patterns of visitation to points of interest, we use Non-negative Matrix Factorization (NMF), a machine learning dimensionality reduction technique, to identify the key behavioral components which describe individuals' mobility activity. This is analogous to identifying the fundamental "genes" which, when taken together, form an individual's unique lifestyle. After decomposing individual-level mobility data into behavioral components, we compute contact networks between individuals based on their proximity in space and time. We then compare observed patterns of interaction with respect to the distribution of individuals' latent behavioral features. Our approach can highlight behavioral homophily in networks, were contacts occur more frequently between individuals with similar behavioral characteristics, and behavior-based community structure which is not captured by demographic features.

Changing technology fosters higher levels of human mobility creating both new connections between faraway places and rearranging the spatial structure of already existing connections. Both mechanisms and their impact on network topology are little understood in the literature.In this work, we address this gap by leveraging a unique longitudinal population-scale network dataset sourced from Statistics Netherlands. This network contains family, work, school, household, and next-door neighbor connections derived from administrative registers, that together constitute a multilayer social opportunity structure for all residents of the Netherlands between 2009 and 2022. We follow the patterns of individuals’ ego networks over time, and measure their size, closure, and geographical dispersion.First, we show that while the average size of ego networks stays stable over the observed period, average closure drops by as much as 10%, leading to more open local network structures. Second, we see that the average geographical distance from network neighbors grows, and in parallel, the average share of network neighbors in the same municipality decreases. Third, we link the two using multivariate difference-in-difference type regressions which show that the observed decrease in the closure is indeed significantly linked to the growing geographic dispersion. The regressions thus confirm that people’s mobility is linked to more open ego networks, which potentially impacts on people’s access to information, or the level of trust in communities.This work is among the first ones that aims to map the temporal network of an entire population structure comprehensively [2]. As such, it offers a starting point for a wide variety of impactful network science research at the level of a complete population.

The phenomenon of gentrification of an urban area is characterized by the displacement of lower-income residents due to rising living costs and an influx of wealthier individuals. This study presents an agent-based model that simulates urban gentrification through the relocation of three income groups -- low, middle, and high -- driven by living costs. The model incorporates economic and sociological theories to generate realistic neighborhood transition patterns. We introduce a temporal network-based measure to track the outflow of low-income residents and the inflow of middle- and high-income residents over time. Our experiments reveal that high-income residents trigger gentrification and that our network-based measure consistently detects gentrification patterns earlier than traditional count-based methods, potentially serving as an early detection tool in real-world scenarios. Moreover, the analysis highlights how city density promotes gentrification. This framework offers valuable insights for understanding gentrification dynamics and informing urban planning and policy decisions.