PaperlayerView exact revision on arXiv

Source-linked AI summary

A Machine Learning Approach to Modeling Human Migration

Caleb Robinson, Bistra Dilkina

arXiv:1711.05462v1cs.SIcs.LGphysics.soc-ph

TL;DR

Migration modeling matters for policy, yet explicit research on migration has received less attention than commuting-flow modeling. This paper develops machine-learning models and a training pipeline for migration flows, reporting better performance than traditional models across US-county and international datasets.

  • Problem

    Migration-flow modeling has received less attention than commuting-flow modeling, despite its relevance to population change, labor markets, disease spread, and trade.

    Source: §1 · p. 1

  • Method

    The paper formulates migration-flow prediction as a machine-learning task and develops a training pipeline that handles imbalance, tuning, evaluation, and a task-suited neural-network loss.

    Source: §1 · p. 1

  • Results

    Machine-learning models outperform traditional human-migration models in most evaluation metrics across US-county and global migration datasets.

    Source: §6 · p. 6

  • Takeaways & Limitations

    Machine-learning models offer flexible ways to combine socioeconomic features nonlinearly and customize migration prediction to a country or problem.

    Source: §6 · p. 6

    Further study of migration under extreme weather requires higher-resolution migration data in space and time, with local scales important because such events are short-lived.

    Source: §6 · pp. 6–7

Abstract

from arXiv · show

Human migration is a type of human mobility, where a trip involves a person moving with the intention of changing their home location. Predicting human migration as accurately as possible is important in city planning applications, international trade, spread of infectious diseases, conservation planning, and public policy development. Traditional human mobility models, such as gravity models or the more recent radiation model, predict human mobility flows based on population and distance features only. These models have been validated on commuting flows, a different type of human mobility, and are mainly used in modeling scenarios where large amounts of prior ground truth mobility data are not available. One downside of these models is that they have a fixed form and are therefore not able to capture more complicated migration dynamics. We propose machine learning models that are able to incorporate any number of exogenous features, to predict origin/destination human migration flows. Our machine learning models outperform traditional human mobility models on a variety of evaluation metrics, both in the task of predicting migrations between US counties as well as international migrations. In general, predictive machine learning models of human migration will provide a flexible base with which to model human migration under different what-if conditions, such as potential sea level rise or population growth scenarios.

1 Introduction

Migration modeling matters for policy, yet research has focused more on commuting flows than on migration. The paper introduces machine-learning approaches and a training pipeline for predicting migration flows.

1 Introduction

  • Migration estimates can inform population change, labor markets, infectious-disease spread, and international trade, while explicit migration modeling has received less attention than commuting-flow modeling.Source: §1 · p. 1
  • Gravity and radiation models use crafted functional forms and limited inputs, trading predictive power for interpretability.Source: §1 · p. 1
  • The paper develops the first general machine-learning formulation of human migration prediction.Source: §1 · p. 1
  • The proposed training pipeline addresses dataset imbalance, hyperparameter tuning, and performance evaluation.Source: §1 · p. 1
  • The paper develops a custom artificial-neural-network loss function suited to migration prediction.Source: §1 · p. 1

4. We compare the performance of machine learning models

Machine-learning models outperform traditional human-migration models across the two datasets evaluated.

4. We compare the performance of machine learning models

  • Machine-learning models outperform traditional human-migration models in all cases across two datasets.Source: §4 · p. 2

2 Traditional Migration Models

Traditional migration models estimate flows by combining origin departures with destination probabilities, using population, distance, and intervening-opportunity information.

2 Traditional Migration Models

  • Traditional models decompose predicted migration from origin i to destination j into outgoing migrants G_i and destination probability P_ij.Source: §2 · p. 2
  • When prior outgoing-migrant counts are unavailable, a production function estimates them as a fitted fraction α of zone population.Source: §2 · p. 2
  • The evaluated radiation, extended-radiation, and gravity models use zone populations and distance, with intervening population also used by radiation models.Source: §2 · p. 2
  • Traditional models are interpretable but depend on tuned parameters, and the paper contrasts that advantage with machine-learning predictive accuracy.Source: §2 · p. 2

3 Evaluation Methods

The evaluation compares predicted and observed migration flows using overlap, distance-distribution, and regression measures, and also assesses incoming-migrant estimates by zone.

3 Evaluation Methods

  • CPC directly compares predicted and ground-truth traveler matrices, ranging from no shared entries to identical matrices.Source: §3 · p. 2
  • CPCd measures whether predicted flows reproduce the ground-truth distribution of trips across distance bins.Source: §3 · p. 2
  • RMSE penalizes larger prediction errors more than smaller ones and can increase without bound as predictions worsen.Source: §3 · p. 2
  • The study also evaluates incoming migrants per zone using MAE and r2, because these estimates inform population-growth, infrastructure, and job analyses.Source: §3 · p. 2

4 Learning Migration Models

The models predict origin–destination migration flows from zone and pairwise features, with a static-in-time formulation. The training pipeline addresses sparse migration data and tunes XGBoost and ANN models, including a CPC-based ANN loss that performs better than MSE on validation performance and training time.

4 Learning Migration Models

  • The formulation predicts each origin-to-destination flow directly from origin, destination, and pairwise features, assuming flows depend on features from the previous timestep.It does not require the production function used by traditional models; the authors note that the formulation can be used where only one year of data is available.Source: §4 · p. 3

Hyperparameter optimization

  • The study tunes XGBoost and ANN models, including tree depth, estimators, and learning rate for XGBoost, and loss, layers, width, epochs, and batch size for ANN.XGBoost also provides a ranking of relative feature importances.Source: Hyperparameter optimization · p. 3

Dealing with zero-inflated data

  • The training pipeline undersamples zero-migration pairs, retaining all observed migration pairs and adding k randomly selected zero-migration pairs per observed pair.The authors note that this necessarily discards available information; k is included in the model parameter searches.Source: Dealing with zero-inflated data · p. 3

Custom ANN loss function

  • The custom ANN loss is derived from CPC because migration counts are zero-inflated, long-tailed, and CPC is a key evaluation metric.MAE and MAPE performed poorly in preliminary experiments, partly because they do not handle large errors and many zeros well, respectively.Source: Custom ANN loss function · p. 3
  • The CPC-based ANN loss yields better validation performance and faster training than MSE loss.It rewards matching migrant counts while also minimizing per-link absolute error; training uses mini-batches, so it is not exactly equivalent to whole-matrix CPC.Source: Custom ANN loss function · p. 4

5 Experiments

Experiments compare machine-learning and traditional models on US county and international migration data, using feature-rich origin–destination observations and year-based validation. Machine-learning models outperform traditional baselines across the reported datasets, while feature rankings and county maps reveal where added features matter.

Datasets

  • USA data cover 3,106 counties from 2004–2014, while global data cover 207 countries across five decadal timesteps from 1960–2000.The datasets combine migration flows with per-zone and between-zone features; the USA set includes 7 per-county and 6 between-county features, and the global set includes 5 per-country and 3 between-country features.Source: Datasets · p. 4

Experimental Setup

  • Year triplets assign consecutive observations to training, validation, and testing, with randomized search over 50 hyperparameter samples and CPC-based selection.Final results average test performance across triplets.Source: Experimental Setup · pp. 4, 6
  • Training downsampling retains every positive-flow observation and samples k times as many zero-flow observations, while validation and testing use full datasets.The tested k ranges differ by dataset: 5–100 for USA and 1–5 for Global, reflecting their different nonzero-flow prevalence.Source: Experimental Setup · p. 4

Results

  • Machine-learning models outperform traditional baselines across the reported USA and Global Migration results, including all metrics on the global dataset.With traditional features and a production function, ANN beats traditional models on 5 of 6 USA measures and XGBoost on 4 of 6; without production functions, both ML models outperform their constrained versions on 5 of 6 metrics.Source: Table 2 · p. 5, Table 3 · p. 5, Results · p. 6
  • Intervening population ranks among the four most important features in both datasets; intervening county count leads in USA, while origin population growth leads globally on average.Global origin population growth importance varies substantially across years and is less predictive in relatively stable years.Source: Results · p. 6
  • County maps show extended-feature ML models correct western underprediction, while the extended-feature ANN better matches rural incoming-migrant counts than traditional models.Traditional models consistently overestimate migrants to rural areas.Source: Results · p. 6

6 Conclusion

The paper presents machine-learning models as flexible alternatives to traditional migration models, reporting better performance on most evaluation metrics. It also identifies limits in explaining migration dynamics and in applying the models to extreme-weather scenarios.

6 Conclusion

  • Machine-learning models can combine many input features nonlinearly and be customized to the problem or country, unlike traditional models with fixed forms.Source: §6 · p. 6
  • The two machine-learning models predict US county and international migration flows and outperform traditional models on most evaluation metrics.Source: §6 · p. 6
  • XGBoost feature-importance rankings do not fully explain migration dynamics, and the authors propose studying other models and extreme-weather migration.Source: §6 · p. 6
  • The 2014 county maps compare ground-truth and predicted incoming migrants for six traditional and machine-learning models, with colors indicating over- or underestimation.Source: Figure 1 · p. 7
  • Higher-resolution migration data across spatial and temporal scales are needed for extreme-weather applications, whose effects are better estimated and predicted locally.Source: §6 · p. 7

Related papers

Loading 1711.05462v1…