PaperlayerView exact revision on arXiv

Source-linked AI summary

A Survey on Deep Learning for Human Mobility

Massimiliano Luca, Gianni Barlacchi, Bruno Lepri, Luca Pappalardo

arXiv:2012.02825v2cs.LGcs.AIcs.SI

TL;DR

Human mobility has broad societal importance, but existing surveys did not comprehensively describe deep-learning solutions across mobility tasks. This survey builds a taxonomy and synthesizes deep-learning approaches, challenges, datasets, and evaluation practices for prediction and generation. It concludes that predictive tasks are well established, while the surveyed solutions retain important limitations and open challenges.

  • Problem

    Existing surveys focus on individual tasks, data sources, or traditional approaches, leaving comprehensive coverage of deep-learning solutions for human mobility missing despite mobility’s societal importance.

  • Method

    The survey organizes and reviews deep-learning solutions for next-location prediction, crowd flow prediction, trajectory generation, and flow generation, including challenges, datasets, metrics, and model modules.

  • Results

    Predictive tasks are well established and addressed by a significant variety of deep-learning approaches, while generative-task adoption is contrasted in the survey’s conclusions.

  • Takeaways & Limitations

    The survey serves as a guide to leading deep-learning mobility solutions and identifies concepts and open challenges for researchers and practitioners.

  • Takeaways & Limitations

    Deep-learning mobility solutions suffer from several limitations, including insufficient investigation of privacy risks and the trade-off between perturbing mobility data and preserving realism or predictive accuracy.

Abstract

from arXiv · show

The study of human mobility is crucial due to its impact on several aspects of our society, such as disease spreading, urban planning, well-being, pollution, and more. The proliferation of digital mobility data, such as phone records, GPS traces, and social media posts, combined with the predictive power of artificial intelligence, triggered the application of deep learning to human mobility. Existing surveys focus on single tasks, data sources, mechanistic or traditional machine learning approaches, while a comprehensive description of deep learning solutions is missing. This survey provides a taxonomy of mobility tasks, a discussion on the challenges related to each task and how deep learning may overcome the limitations of traditional models, a description of the most relevant solutions to the mobility tasks described above and the relevant challenges for the future. Our survey is a guide to the leading deep learning solutions to next-location prediction, crowd flow prediction, trajectory generation, and flow generation. At the same time, it helps deep learning scientists and practitioners understand the fundamental concepts and the open challenges of the study of human mobility.

1 INTRODUCTION

Human mobility matters across public health, urban systems, transportation, and environmental outcomes, while digital traces and deep learning enable increasingly comprehensive study. This survey organizes deep-learning research around four mobility tasks, reviews their solutions and datasets, and identifies open challenges.

  • Human mobility affects disease spreading, disaster response, transportation, citizen well-being, pollution, and resource consumption.
  • Digital mobility data from GPS devices, vehicles, boats, cellular networks, and social media provide movement observations at multiple temporal and spatial scales.
  • Existing surveys separately address individual tasks, data sources, or traditional machine-learning approaches without specifically focusing on deep-learning mobility challenges and solutions.
  • The survey addresses next-location prediction, crowd flow prediction, trajectory generation, and flow generation through a predictive-versus-generative taxonomy.
  • The survey also discusses open challenges and provides a GitHub repository for maintaining relevant mobility datasets and papers.
  • For each task, the survey defines the problem, reviews traditional approaches, explains deep-learning advantages, and identifies commonly used modules, datasets, and evaluation metrics.

2 BACKGROUND

The background defines mobility data as electronic records of movement, especially trajectories and spatially aggregated flows, and introduces the deep-learning modules used throughout the survey. These concepts establish how individual movements are represented and how models capture spatial, temporal, external, and generative patterns.

  • 2.1 Spatio-temporal trajectories and spatial aggregations: Mobility data describe movements over an observation period and are typically stored as spatio-temporal trajectories or mobility flows.
  • 2.1 Spatio-temporal trajectories and spatial aggregations: A trajectory is an ordered sequence of spatio-temporal records containing an individual identifier, geographic location, and timestamp.
  • 2.1 Spatio-temporal trajectories and spatial aggregations: Spatial tessellation partitions an area into finite, non-overlapping tiles whose union covers the area, enabling mobility points or flows to be assigned spatially.
  • 2.2 Deep-learning modules: Fully connected networks capture effects of external features and preferences, while RNNs, LSTMs, and GRUs model sequential spatial and temporal mobility patterns.
  • 2.2 Deep-learning modules: Attention mechanisms emphasize contextually important inputs, and CNNs capture spatial patterns, particularly in crowd flow prediction.
  • 2.2 Deep-learning modules: VAEs and GANs generate realistic synthetic trajectories by encoding data into latent distributions or training generators against discriminators.

3 PREDICTIVE MODELS

Predictive models forecast future whereabouts at individual or collective levels, with deep learning addressing spatial, temporal, and external-factor dependencies across mobility tasks.

  • Predictive models forecast individuals’ future locations or aggregated crowd flows from historical mobility observations.
  • Next-Location Prediction: Next-location prediction forecasts an individual’s next stay point and supports applications including public-health monitoring, transportation planning, travel recommendation, and geomarketing.
  • Next-Location Prediction: Next-location predictors must capture spatial and temporal regularities, departures from routine, external factors, and individual preferences such as weather, points of interest, and friendships.
  • Next-Location Prediction: Selected next-location studies use check-in and GPS trajectory datasets, including Foursquare, Gowalla, Porto taxi traces, and San Francisco taxi traces, with distance or classification metrics.
  • Next-Location Prediction: Deep-learning solutions include fully connected, spatial-temporal recurrent, and attentional recurrent models, with DeepMove combining multimodal embeddings, attention, and GRU processing.
  • Crowd Flow Prediction: Crowd flow prediction forecasts incoming and outgoing flows across spatially tessellated regions, requiring models to represent spatial dependencies and temporal closeness, trends, and periodicity.

4 GENERATIVE MODELS

Generative mobility models produce synthetic individual trajectories or mobility flows that reproduce observed spatial, temporal, social, and geographic patterns. The survey formalizes these tasks, reviews their evaluation, and explains how deep learning models address limitations of traditional approaches.

  • Generative models: Generative mobility models create realistic spatio-temporal trajectories or mobility flows.Individual trajectory generation targets synthetic movements, while flow generation estimates flows between location pairs.
  • Trajectory generation: A realistic trajectory generator must capture spatial and temporal patterns, including visited locations, visit timing, durations, routine behavior, and sporadic mobility.The simultaneous modeling of routine and out-of-routine movement is identified as a major challenge.
  • Evaluation: Generative models are evaluated by comparing distributions of mobility patterns computed on synthetic trajectories with those from held-out real trajectories.Lower KL or JS divergence indicates greater realism for the corresponding mobility measure and test set.
  • Deep learning approaches: Deep learning generative paradigms such as GANs and VAEs can capture spatial, temporal, and social dimensions simultaneously, including complex non-linear relationships.This flexibility allows them to generate more realistic data than traditional approaches that often capture only particular mobility aspects.
  • Datasets and surveyed solutions: Trajectory generators commonly use GPS datasets such as Geolife, MDC, and San Francisco taxi traces, with realism assessed through standard mobility metrics and distribution divergences.The surveyed approaches are predominantly GAN-based, with a smaller number using VAEs.
  • Selected deep learning solutions: The survey reviews trajectory-generation implementations using GANs, CNNs, RNNs, VAEs, and LSTMs across selected datasets and evaluation settings.Examples include Wasserstein GANs with CNN-based generators and discriminators, GANs using matrix representations, and SVAE combining a VAE with an LSTM.

5 CONCLUSIONS

The survey organizes deep-learning research on human mobility around predictive and generative tasks, finding predictive tasks more established while generative tasks remain newer and need further attention. It also identifies open challenges involving transferability, explainability, privacy, tunability, and interactions among individuals.

  • Predictive tasks are well established, whereas deep-learning applications to trajectory and flow generation are more recent and need further attention.
  • Geographic Transferability: Geographic transferability remains open because models depend on training data and may not transfer between distinct, non-overlapping regions.This matters when mobility data are scarce or absent in the target region.
  • Explainability: Deep-learning models are opaque, making it difficult to reconstruct predictions or generated trajectories and identify feature roles or reasoning biases.The survey connects explainability to useful knowledge for policymakers and urban planners.
  • Privacy: Privacy risks arise during training and prediction or generation, including possible re-identification from synthetic trajectories and leakage of private training information.The extent to which mobility data can be perturbed without degrading realism or accuracy remains barely investigated.
  • Tunability: Current models have limited tunability and may not preserve realism or accuracy across different spatial tessellations, mobility restrictions, events, or temporal granularities.Administrative tiles may support interpretability, while fine-grained grids may help analyze transportation networks and major events.
  • Interaction Dimension: Next-location predictors and trajectory generators often assume independent individuals despite collective movements that contribute to traffic jams, accidents, and massive commuting patterns.

B.2 Convolutional Neural Networks

CNNs extract and downsample spatial features through convolution and pooling, typically using a fully connected layer for prediction. In human mobility, they primarily model spatial dependencies in crowd-flow matrices and may be combined with recurrent networks for temporal dependencies.

  • CNNs alternate convolutional layers that extract features with pooling layers that downsample spatial representations and reduce computational requirements.Pooling can use maximum or average statistics and provides approximate invariance to small input translations.
  • A CNN architecture can use two convolutional and two pooling layers followed by a fully connected layer that outputs the prediction.
  • Vanishing gradients limit deep CNNs, while residual units address the issue through skip connections that preserve information across layers.The skip size is usually smaller than 4.
  • CNNs are used mainly for crowd-flow prediction, representing incoming and outgoing flows as sequences of matrices to capture spatial dependencies among city areas.They are often combined with RNNs to capture temporal dependencies.

C DATA

Digital traces from GPS devices, mobile-phone networks, and social media provide detailed observations of human movement. Mobile-phone data can cover large populations but have coarse spatial resolution, privacy restrictions, and reproducibility challenges.

  • Digital mobility datasets include GPS tracks, mobile-phone records, and geotagged social-media posts, but most massive datasets are proprietary and difficult to reproduce.
  • Call Detail Records encode users, timestamps, communicating radio base stations, and call duration, enabling mobility reconstruction from consecutive radio-base-station records.
  • XDRs record a user, timestamp, serving radio base station, and transferred-data amount, and can partially reduce the sparsity found in CDRs.
  • Mobile-phone positions are approximated at radio-base-station level, leaving within-coverage location unknown while enabling potentially national-scale samples.Sensitive information means these data are typically unavailable for public sharing, limiting reproducibility.

C.1.1 Available datasets.

GNSS provides geographic positioning through satellite signals, with GPS widely used in phones, vehicles, vessels, and wearable devices. GeoLife is a public GPS-trajectory dataset covering 182 users over 4.5 years at fine temporal or spatial sampling.

  • GNSS determines longitude, latitude, altitude, and time from satellite signals, while GPS is its most popular system.
  • GPS receivers are embedded in mobile phones, vehicles, vessels, and wearable devices, with phones activating GPS through location-dependent applications.
  • GeoLife publicly provides GPS trajectories from 182 users collected over 4.5 years at intervals of 1–5 seconds or 5–10 meters.

C.2.1 Available datasets.

The survey identifies several public mobility datasets spanning taxi, bike-sharing, GPS, and pandemic-related mobility-flow data. These datasets capture trajectories, station trips, spatial flows, density, and related mobility attributes across cities and regions.

  • Public taxi datasets provide trip trajectories with geographic, temporal, and vehicle-related attributes across multiple cities.San Francisco trajectories are sampled about every 10 seconds and include taxi identifiers, coordinates, timestamps, and occupancy; Porto records contain coordinates and timestamps.
  • The T-Drive dataset contains one week of trajectories from about 10,000 Beijing taxis, sampled every 177 seconds.Each point includes the taxi identifier, latitude, longitude, and timestamp.
  • Public flow datasets represent incoming and outgoing mobility across spatial tiles in New York, Beijing, Tokyo, and Osaka.The sources include taxi GPS, bike-sharing systems, and GPS devices; the Tokyo and Osaka dataset also provides people density.
  • Bike-sharing datasets record trips between stations, including origin and destination stations, coordinates, and trip times.Citi Bike covers New York City, with similar open datasets available for other cities such as Washington DC.
  • The Mobile Data Challenge dataset contains trajectories and multimodal behavioral information from 185 participants in Lausanne.Its mobility files use GPS receivers and wireless local area networks, alongside calendar and accelerometer data.
  • During the COVID-19 pandemic, Cuebiq and SafeGraph provided free data access upon request, including aggregated mobility flows at multiple spatial scales.SafeGraph-based data include country-to-country, state-to-state, and census-tract-to-census-tract flows.

C.3 Social Media Data

Social media posts can provide spatially and temporally tagged observations from which user trajectories may be reconstructed. Their usefulness is balanced by API restrictions and sparse location observations.

  • Geotagged social media posts associate user-generated content with a location and time, enabling reconstruction of trajectories from sequential posts.Platforms may provide precise coordinates or positions associated with predefined locations such as cities, areas, or venues.
  • A geotagged record generally includes user and resource identifiers, posting time, and platform-dependent venue or location information.Location may appear as a venue identifier, category, textual string, or latitude/longitude pair.
  • Social media data are often downloadable through APIs, but platforms impose query limits or require user authorization.These restrictions constrain access to posts and may complicate data collection.
  • Because users’ locations are recorded only when they post or check in, social media mobility data suffer from sparsity.The data nevertheless offer an objective definition of location that facilitates data preparation.
  • Inactive platforms such as Gowalla and Brightkite provide freely available check-in datasets.Gowalla records more than six million check-ins over approximately one and a half years.

C.3.1 Available datasets.

Available social-media mobility datasets include check-ins and geotagged tweets, with coverage varying by platform, city, time span, and representation of location. The survey’s dataset table organizes these resources by scope and mobility task.

  • A Foursquare dataset contains check-ins from 16,000 users over one year in New York City.
  • Twitter mobility datasets commonly represent locations as semantic points of interest or latitude/longitude pairs.The GMove dataset contains 1.4 million tweets from August through November 2014 covering Los Angeles.
  • A public repository provides a comprehensive list of Twitter datasets.
  • SafeGraph-based data provide aggregated daily mobility flows across country, state, and census-tract spatial levels.
  • The survey’s dataset table records each dataset’s reference, item count, time span, geographic coverage, selected papers, mobility tasks, and download link.A dash indicates that the dataset is aggregated, unavailable in count, or continuously updated.

D.1 Distance metrics

The survey uses geographic distance metrics to compare locations on Earth. The Haversine distance is defined for two points on the spherical Earth, with lower values indicating better performance.

  • The Haversine distance measures the spherical-Earth distance between two points using their longitudes, latitudes, and the Earth’s radius.The longitude and latitude variables correspond to the two input points, while R denotes the Earth radius.
  • Haversine distance ranges from 0 to infinity, and lower values indicate better performance.

D.2 Classification metrics

The survey defines metrics for evaluating next-location classifiers, generative mobility models, and trajectory similarity, then reviews statistical regularities used to characterize human mobility.

  • Classification metrics: ACC measures correctly predicted locations, while ACC@k counts cases where the true location appears among the k most likely predictions.ACC equals ACC@1.
  • Classification metrics: Precision, recall, and F1-score evaluate positive-class prediction using true positives, false positives, and false negatives.F1-score is the harmonic mean of precision and recall.
  • Classification metrics: ROC curves plot recall against false positive rate, while AUC summarizes classifier performance across thresholds and ranges from 0 to 1.AUC is scale-invariant, and AUC=0.50 represents random prediction performance.
  • Error metrics: MAE, MSE, RMSE, and MAPE measure prediction errors, with lower values indicating better performance.MSE and RMSE penalize large errors more strongly than MAE, while MSE is sensitive to outliers.
  • Generative-model metrics: CPC measures similarity between generated and real flows, with 1 indicating a perfect match and 0 indicating no overlap.The Sørensen-Dice index is also called the Common Part of Commuters measure.
  • Spatial patterns: Mobility patterns include ranked visitation frequencies, movement distances, mobility entropy, network motifs, and characteristic travel radii.The survey describes Zipf-like location visitation, truncated power-law displacement and radius distributions, and returner–explorer differences.

E.2 Temporal metrics

Temporal mobility metrics describe how long people remain between recorded locations and how visits vary by time and stay duration.

  • Temporal metrics: Waiting time is the elapsed time between consecutive trajectory points, and its distribution is well approximated by a truncated power-law.Individuals also follow circadian rhythms, tending to remain stationary during night hours.
  • Temporal metrics: Temporal popularity measures the probability of visiting a location at a given time, while staying patterns measure visit probability for a given duration.These metrics distinguish when locations are visited from how long visits last.
Loading 2012.02825v2…