Source-linked AI summary
Universal Predictability of Mobility Patterns in Cities
Xiao-Yong Yan, Chen Zhao, Ying Fan, Zengru Di, Wen-Xu Wang
TL;DR
City mobility prediction needs an accurate approach that does not depend on extensive prior traffic measurements. The paper develops a parameter-free population-weighted opportunities model using population distribution, and reports agreement with observed city mobility patterns across distance distributions, destination constraints, and flux. Its scope is city-scale prediction, where the radiation model is described as underestimating mobility.
Problem
Existing mobility models often rely on fitted traffic data, while models successful at larger spatial scales can underestimate relatively high mobility within cities.
Method
The paper develops a population-weighted opportunities model without adjustable parameters, using population distribution and competition among destinations to predict city mobility.
Results
The model agrees with real data for travel distance distributions, destination travel constraints, and flux across cities with different sizes, economic levels, and cultural backgrounds.
Takeaways & Limitations
Population distribution alone can support city-scale mobility prediction while distinguishing city mobility from patterns modeled successfully at larger spatial scales.
Takeaways & Limitations
The radiation model remains unable to provide satisfactory city-scale predictions because it underestimates relatively high urban mobility.
Abstract
from arXiv · showhide
Despite the long history of modelling human mobility, we continue to lack a highly accurate approach with low data requirements for predicting mobility patterns in cities. Here, we present a population-weighted opportunities model without any adjustable parameters to capture the underlying driving force accounting for human mobility patterns at the city scale. We use various mobility data collected from a number of cities with different characteristics to demonstrate the predictive power of our model. We find that insofar as the spatial distribution of population is available, our model offers universal prediction of mobility patterns in good agreement with real observations, including distance distribution, destination travel constraints and flux. In contrast, the models that succeed in modelling mobility patterns in countries are not applicable in cities, which suggests that there is a diversity of human mobility at different spatial scales. Our model has potential applications in many fields relevant to mobility behaviour in cities, without relying on previous mobility measurements.
I. INTRODUCTION
City-scale mobility prediction remains difficult because established models often require fitted traffic data or underestimate residents’ relatively high mobility. The paper introduces a parameter-free population-weighted opportunities model that uses population distribution to represent destination competition across the city.
- Motivation: Gravity and related trip-distribution models require context-specific parameters fitted from mobility data, limiting their use where previous measurements are unavailable.This limitation motivates models that can predict mobility from lower data requirements.
- Motivation: The radiation model uses only population distribution and no adjustable parameters, but evidence indicates that it may not predict mobility adequately at the city scale.Its success at larger spatial scales does not guarantee applicability within cities.
- Motivation: Higher urban mobility allows residents to travel farther toward locations with greater opportunities, challenging models developed for larger spatial scales.The paper links this city-scale setting to the need for a different mobility mechanism.
- Evaluation: The paper evaluates city-scale mobility using destination-selection patterns and travel-distance distributions across multiple cities.The figure descriptions compare model-generated patterns with observed data and define Pdist(r) as the probability of travel at distance r.
- Population-weighted opportunities model: The PWO model treats destination attraction as proportional to opportunities and inversely weighted by population in the area between origin and destination.Population represents opportunities, while intervening population represents competition among potential destinations.
- Population-weighted opportunities model: The model enlarges the possible destination area to the whole city while retaining distance-related decay in travel probability.This design addresses relatively high city-scale mobility without introducing adjustable distance parameters.
B. Predicting mobility patterns
The PWO model is validated against real mobility data across cities using distance distributions, destination travel constraints, and travel fluxes. Across these measures, it generally agrees better with observations than the radiation model.
- PWO-predicted travel distance distributions agree well with real distributions, while the radiation model underestimates travel beyond approximately 2 km.The comparison covers daily travel data from four cities collected using GPS, mobile phones, and household surveys.
- Pdest(m), the probability of traveling to a location with population m, is predicted equally or better by PWO than by the radiation model.Pdest(m) evaluates destination travel constraints for origin-constrained mobility models.
- PWO flux predictions are in reasonable agreement with real observations, whereas radiation-model averages deviate except in Abidjan.Fluxes are evaluated between all pairs of locations; the figure compares predicted and observed travel counts using binned averages and boxplots.
- The Sørensen similarity index provides a clearer comparison than the boxplot method because it quantifies similarity with real observations.The boxplot method cannot explicitly distinguish the two models’ performance in every case.
III. DISCUSSION
The discussion presents PWO as a parameter-free, population-based model that predicts city-scale mobility across diverse cities, while clarifying both the radiation model’s scale limitation and PWO’s remaining accuracy gap. Its travel matrices share approximately 70% of their common part with real data.
- III. DISCUSSION: PWO uses only spatial population distribution, without adjustable parameters, to reproduce city mobility across differences in size, economic level, and cultural background.The model agrees with real data for travel distance distribution, destination travel constraints, and flux.
- III. DISCUSSION: The radiation model performs well at large spatial scales but cannot satisfactorily predict city-scale mobility because it underestimates relatively high urban mobility.Its assumption that limited mobility prevents farther destination choices is described as reasonable between cities but inappropriate within cities.
- III. DISCUSSION: PWO overcomes this city-scale limitation by assuming destination attraction is inversely proportional to population, reflecting competition for opportunities across the whole city.This assumption is presented as the basis for improved city-scale prediction when population distribution is available.
- III. DISCUSSION: Parameterized gravity, intervening-opportunity, and rank-based models can occasionally be more accurate, but their parameter dependence limits their scope to cases with particular previous information.In PWO and the radiation model, the decrement is determined naturally by population distribution rather than adjustable parameters.
- III. DISCUSSION: Approximately 70% of PWO travel-matrix structure is common with real data, so predictability remains below the average upper limit of human mobility.The authors attribute remaining inaccuracy partly to aggregate models’ inability to represent individual behavioral diversity; microscopic models may improve prediction but have higher computational complexity.
IV. MATERIALS AND METHODS
The study evaluates mobility prediction across Beijing, Shenzhen, Abidjan, and Chicago using taxi, mobile-phone, and travel-survey data. These datasets cover different urban settings and provide observed movements for comparison.
- Data sets: Beijing and Shenzhen datasets record taxi-passenger trips, while Abidjan data track movements among cell-phone antennas.The Beijing dataset contains 1,070,198 records, and the Shenzhen dataset contains 2,338,576 trips; Abidjan data contain 607,167 users’ movements.
- Data sets: The four-city dataset combines automatically recorded taxi trajectories, anonymised mobile-phone movements, and household travel diaries.These sources differ in measurement method and sampling population, enabling comparisons across urban datasets.
- Data sets: Chicago mobility is represented by a travel-tracker survey covering 10,552 households in 2007–2008.The survey provides household members’ travel inventories, including trip origins and destinations.
B. Data preprocessing
Raw origin–destination coordinates are converted into zone-level mobility data, while population distributions support radiation-model predictions. Trips are then aggregated into origin totals and origin–destination fluxes.
- Data preprocessing: Zone-level totals include Ti, the trips departing from zone i, and Tij, the trips traveling from zone i to zone j.The resulting aggregates form the travel quantities used by the mobility models.
- Data preprocessing: The radiation model predicts travel fluxes among locations using population distribution without adjustable parameters.Its formulation uses origin population mi, destination population mj, intervening population sij, and total departures Ti.
- Data preprocessing: The radiation-model flux expression incorporates the populations at the origin and destination together with intervening population within the origin-centered distance circle.The intervening population excludes the origin and destination zones.
D. Sørensen similarity index
The Sørensen similarity index measures how closely predicted and observed travel fluxes agree. It ranges from 0 for strongly mismatched fluxes to 1 for exact agreement.
- D. Sørensen similarity index: The index compares predicted trips T′ij with observed trips Tij using twice their pairwise minimum, aggregated across fluxes.It is used to assess whether mobility models reproduce real fluxes on average.
- D. Sørensen similarity index: An index close to 0 indicates that predicted fluxes are far from the observed values.This represents poor agreement between modelled and real travel flows.
- D. Sørensen similarity index: An index of 1 indicates that predicted and observed fluxes are equal.The maximum value corresponds to exact reproduction of the observed travel flows.
A. Data collection and preprocessing
The paper assembles mobility observations from European and U.S. cities, preprocesses them into spatial zones, and compares parameterised and parameter-free models. The methods include the PWO, radiation, gravity, intervening-opportunity, and rank-based approaches.
- A. Data collection and preprocessing: Gowalla check-ins provide 6,442,890 records from February 2009 to October 2010 for evaluating European-city mobility.A user trip is defined by two consecutive check-ins at different locations, with trips retained when endpoints fall within a city.
- B. Prediction results: The PWO model and radiation model are compared using travel-distance distributions and fluxes between all location pairs.For European cities, PWO predictions agree with observations whereas radiation-model results deviate.
- A. Data collection and preprocessing: European-city analyses use zone partitions and population-density distributions for London, Berlin, Prague, Oslo, Copenhagen, and Goteborg.These inputs support comparisons of travel-distance distributions and travel fluxes.
- A. Data collection and preprocessing: U.S.-city travel surveys provide origin–destination records for New York, Seattle, Detroit, and Twin Cities after coordinate-based processing.The surveys contain household and trip information, while the resulting zones and population distributions are shown in Figure S4.
- A. The parameterised models: The gravity, intervening-opportunity, and rank-based models use distance functions, opportunities, or rank-distance formulations, generally with adjustable parameters.The gravity model may use a power distance function, while the rank-based model includes adjustable parameter γ.
- A. The parameterised models: The intervening-opportunity model treats destination choice as depending on opportunities and intervening population rather than distance alone.Its decision process assumes a constant probability of satisfaction at each opportunity and uses population as a proxy for opportunities.
- A. The parameterised models: The rank-based model makes destination choice depend inversely on rank-distance, where rank 1 denotes the closest location and γ is adjustable.Its formulation also includes origin and destination populations and inter-location distance.
- A. The parameterised models: The rank-based model can violate the origin constraint Ti = Σj Tij, motivating constrained gravity-model alternatives.Doubly constrained gravity models require iterative balancing factors, while singly constrained versions simplify calculation.
B. Estimating model parameters
Parameterized mobility models are calibrated by matching their modeled average travel distance to the real average distance, using the Hyman method and secant iterations.
- The Hyman method estimates model parameters by minimizing the difference between modeled and real average travel distances.
- The error E(β) compares the model’s average distance with the real average travel distance using modeled and observed trip matrices.
- The secant procedure starts with β0 = 1/¯r, computes a trip matrix, and updates β until modeled and real average distances are sufficiently close.
- Estimated parameters for the gravity, I. O., and rank-based models are listed for different cities in Table S3.
C. Comparison among different models
Across travel-distance distributions, destination constraints, and travel fluxes, the parameter-free PWO model generally agrees well with observations and often outperforms alternative models, while parameterized gravity models can sometimes do better.
- Travel distance distribution: Gravity and rank-based models reproduce observed travel-distance distributions in most cases, whereas the I. O. model shows significant deviations.
- Supplementary comparisons examine power-law and exponential gravity distance functions alongside predicted and observed fluxes.
- Unlike gravity models requiring parameters estimated from previous mobility measurements, the PWO model uses only population distribution as input.
- Destination travel constraints: The PWO model matches destination travel constraints at least as well as the other models in all cases.
- Travel fluxes between all pairs of locations: Predicted average fluxes from the PWO, gravity, and rank-based models are comparable with real fluxes to some extent across Asian, African, and U. S. cities.
- Travel fluxes between all pairs of locations: On average, the PWO model has higher Sørensen similarity index accuracy than the I. O. and rank-based models, although gravity can perform better in some cases.
S4. RELATIONSHIP AMONG TRIP DISTRIBUTION MODELS
The paper relates several trip-distribution models by showing that, under uniform population, they can become gravity-like and fall into two selection-framework categories.
- Under uniform population distributions, the PWO, radiation, I. O., and rank-based models can all transform into gravity-like models.
- These models are classified into sequential-selection and global-selection modeling frameworks.
A. Uniform population distribution
Under uniform population, the models can be expressed in gravity-like forms, revealing how distance, destination population, and selection assumptions shape their predictions.
- For the PWO model, a uniform population distribution yields a gravity distance function f(rij) = r^-2 with a cutoff determined by city area A.
- Because real-city populations are nonuniform, this uniform-distance function cannot be directly used without estimating parameters from real traffic data.
- The PWO model instead captures population heterogeneity through a destination population function, with Figure S14(B) showing agreement between destination population and travel proportion in Abidjan.
- Under uniform population, the radiation model becomes a gravity model with power exponent β = 4, making its selection scope more local than the PWO model’s β = 2 form.
- City gravity-law exponents range from 1.63 −2.43, closer to the PWO uniform exponent than to the radiation-model exponent.
- The I. O. model’s high-order exponential distance function explains its usual underestimation of long-distance travel.
- The rank-based model can resemble gravity and PWO distance distributions but ignores destination population, producing inaccurate destination travel constraints.
- All four models share a mechanism in which destination choice decreases as prohibitive factors increase, including distance or rank-distance.
B. Sequential selection and global selection
The paper distinguishes sequential selection, where travellers rank destinations and consider them in order, from global selection, where they evaluate all destinations simultaneously. Different sequential-selection probabilities recover established models, while global selection includes gravity, PWO, and rank-based models.
- Sequential selection: Sequential selection ranks potential destinations by increasing distance and considers them one at a time.The traveller first decides whether to select the nearest-ranked destination before proceeding to others.
- Sequential selection: Assuming θj is proportional to destination population yields the I. O. model.The resulting expression gives the probability of selecting destination j from origin i.
- Sequential selection: Using destination population divided by intervening population produces the radiation model.This assumption is described as the ratio of destination population to total population Sij between locations i and j.
- Sequential selection: Using mj divided by the remaining population M − Sij + mj produces the uniform selection model.The paper presents this as another form of the sequential-selection probability θj.
- Global selection: Global selection evaluates all possible destinations simultaneously and selects among them with probability proportional to destination attraction.The global-selection category includes the gravity, PWO, and rank-based models.
- Comparison: Both frameworks prefer closer destinations, but they encode that preference differently through selection priority or slower attraction decay with distance.The paper compares the frameworks as alternative representations of travellers’ destination decision-making.