PaperlayerView exact revision on arXiv

Source-linked AI summary

Learning Geo-Contextual Embeddings for Commuting Flow Prediction

Zhicheng Liu, Fabio Miranda, Weiting Xiong, Junyan Yang, Qiao Wang, Claudio T. Silva

arXiv:2005.01690v1physics.soc-phcs.LGcs.SI

TL;DR

Urban planning needs commuting-flow predictions from infrastructure and land-use information, but city flows involve complex spatial interactions. GMEL learns geographic-context embeddings with separate supply and demand GATs, then predicts flows; the authors report improved accuracy on New York City data. Its modeling assumes flows are determined by origin supply and destination demand characteristics.

  • Problem

    Urban planning needs commuting-flow predictions from infrastructure and land-use information to understand spatial interactions between city regions.

    Source: Introduction · p. 1

  • Method

    GMEL uses separate graph attention networks to encode geographic context into supply and demand embeddings, trains them with flow and in/out-flow tasks, and predicts flows from the embeddings.

    Source: Introduction · p. 2, Framework · p. 4

  • Results

    The authors report that geographic contextual information greatly improves prediction accuracy and that GMEL is effective against the state of the art on New York City datasets.

    Source: Introduction · p. 2, Conclusion · p. 8

  • Takeaways & Limitations

    Commuting-flow estimates can inform urban planners’ and policy makers’ what-if decisions about new jobs and supporting infrastructure.

    Source: Introduction · p. 2

    GMEL assumes flows are determined by the origin unit’s supply characteristics and the destination unit’s demand characteristics.

    Source: Modeling Supply and Demand… · p. 4

Abstract

from arXiv · show

Predicting commuting flows based on infrastructure and land-use information is critical for urban planning and public policy development. However, it is a challenging task given the complex patterns of commuting flows. Conventional models, such as gravity model, are mainly derived from physics principles and limited by their predictive power in real-world scenarios where many factors need to be considered. Meanwhile, most existing machine learning-based methods ignore the spatial correlations and fail to model the influence of nearby regions. To address these issues, we propose Geo-contextual Multitask Embedding Learner (GMEL), a model that captures the spatial correlations from geographic contextual information for commuting flow prediction. Specifically, we first construct a geo-adjacency network containing the geographic contextual information. Then, an attention mechanism is proposed based on the framework of graph attention network (GAT) to capture the spatial correlations and encode geographic contextual information to embedding space. Two separate GATs are used to model supply and demand characteristics. A multitask learning framework is used to introduce stronger restrictions and enhance the effectiveness of the embedding representation. Finally, a gradient boosting machine is trained based on the learned embeddings to predict commuting flows. We evaluate our model using real-world datasets from New York City and the experimental results demonstrate the effectiveness of our proposal against the state of the art.

Introduction

Commuting flows reflect spatial interactions shaped by infrastructure and land use, but existing approaches may miss complex city patterns or nearby-region influence. GMEL uses geographic context to learn embeddings for flow prediction, and experiments on New York City datasets show effectiveness against the state of the art.

Introduction

  • Commuting-flow predictions can inform urban-planning and transportation-planning decisions, including where to develop districts or deploy new transport modes.The paper gives the example of assessing where workers might commute if a new industrial park is planned.Source: Introduction · pp. 1–2
  • Unlike traffic origin-destination forecasting, commuting-flow prediction estimates edge-level flows from node attributes rather than historical movement features.The paper describes commuting behavior as a daily repeated static pattern.Source: Introduction · p. 2
  • Gravity and radiation models make simple assumptions about flow generation, while machine-learning models may ignore nearby regions.Commuting patterns are also shaped by access to public services, open spaces, transportation, and entertainment.Source: Introduction · pp. 1–2
  • GMEL uses geographic context and graph attention to learn embeddings, then feeds them to a gradient boosting machine for flow prediction.Separate GATs encode supply and demand characteristics, while multitask objectives constrain the embeddings.Source: Introduction · p. 2
  • GMEL experiments on real-world New York City datasets demonstrate effectiveness against the state of the art.The conclusion reports that introducing geographic contextual information can greatly improve prediction accuracy.Source: Introduction · p. 2

Related Work

Commuting-flow models range from gravity-based assumptions to machine-learning regressors, while graph methods offer ways to represent neighboring regions. GMEL draws on geographic context to address the omission of nearby-region influence in commuting-flow prediction.

Commuting Flow Prediction

  • Commuting-flow prediction estimates edge weights from node attributes, unlike traffic OD forecasting, which can use historical network edges as time-series inputs.Source: Commuting Flow Prediction · p. 2
  • Gravity models simplify flow generation; the original form assumes flows scale with origin and destination populations and decay with travel distance.The paper notes that this simple assumption might not capture the complex nature of a city.Source: Commuting Flow Prediction · p. 2
  • Machine-learning regressors can capture complex spatial interactions, but models using only origin-destination attributes ignore nearby regions.GMEL uses geographic context to encode the influence of nearby regions in node embeddings.Source: Commuting Flow Prediction · pp. 2–3

Graph Representation Learning

  • Graph attention aggregates neighbor messages with different weights, and GMEL adapts this mechanism to capture geographic context.Prior graph-learning work includes GraphSAGE and GAT; few studies had used graph neural networks for commuting-flow prediction.Source: Graph Representation Learning · p. 3

Preliminaries

The paper formulates commuting-flow prediction over urban geographic units represented by attributes, geographic relationships, and travel distances. It predicts the flow between an origin and destination from this information.

Preliminaries

  • Urban geographic units can be street blocks, census tracts, or zip-code areas, and this study uses New York City census tracts.The geo-adjacency network illustration represents census-tract centroids as dots and edges as lines.Source: Preliminaries · p. 3, Figure 2 · p. 3
  • Urban indicators are per-unit vectors that aggregate infrastructure and land-use information.Source: Preliminaries · p. 3
  • The geo-adjacency network is an undirected weighted graph whose nodes carry urban indicators and whose edges represent geographic adjacency and relationship strength.Edge features can describe travel distance or trip duration.Source: Preliminaries · p. 3
  • Commuting trips are origin-destination triplets with flow equal to the number of commuters, and the prediction problem uses the geo-adjacency network and distance matrix to estimate these flows.The paper also defines each unit’s total outgoing and incoming commuter flows.Source: Preliminaries · p. 3

Methodologies

GMEL encodes geographic context in separate supply- and demand-oriented embeddings, then trains them with commuting-flow and in/out-flow tasks. A gradient-boosted regression tree uses the learned embeddings to predict flows.

Framework

Graph Attention Network

  • GAT transforms node and edge features into messages, scores edges with attention, normalizes those scores by softmax, and aggregates neighborhood and self impacts.The attention score uses the messages of the connected nodes and their edge features.Source: Graph Attention Network · p. 4

Modeling Supply and Demand Characteristic

  • The model assumes flows depend on origin supply characteristics and destination demand characteristics, using separate GATs whose role-specific attention weights produce distinct embeddings.The two GATs share a structure but assign different weights based on origin or destination roles.Source: Modeling Supply and Demand… · p. 4

Multitask Learning

Training Algorithm

  • After end-to-end GMEL training, a GBRT predicts flows from concatenated origin and destination embeddings plus travel distance.The paper describes GMEL training as pre-training for the flow predictor.Source: Methodologies · p. 4, Training Algorithm · p. 5

Experiments

The evaluation uses New York City census tracts and commuting data, alongside land-use and infrastructure indicators. Trips are split into training, validation, and test sets.

Datasets

  • The evaluation uses 2,168 New York City census tracts as geographic units and randomly splits commuting trips 6:2:2 across training, validation, and test sets.The experimental settings are similar to those reported in Pourebrahim et al. (2019).Source: Datasets · p. 5, LODES · p. 5
  • The 2015 LODES data supplies home-to-work commuting trips aggregated to geographic-unit flows, while OSRM centroid-to-centroid distances serve as geo-adjacency edge features.LODES records workers’ home and employment locations.Source: LODES · p. 5, OSRM · p. 5
  • PLUTO land-use and infrastructure records are aggregated from tax lots into 65 urban indicators for each census tract.The indicator summary is presented in Table 1.Source: LODES · p. 5

Baselines

The comparison includes conventional gravity models, machine-learning predictors, graph embeddings, and GMEL variants. The stated GMEL configuration uses multitask learning and two separate GATs.

Baselines

  • Baselines include power-law and exponential-decay gravity models, random forest, GBRT, and Node2vec embeddings paired with a GBRT flow predictor.The paper describes random forest as reported state of the art in prior work.Source: Baselines · pp. 5–6
  • GMEL-noMul removes the in/out-flow subtasks, while GMEL-noSep replaces the two role-specific GATs with one shared GAT.These variants test the multitask setting and separate-GAT design, respectively.Source: Baselines · p. 6
  • The compared GMEL uses λmain = 0.5 and λsub = 0.5, 128-dimensional embeddings for each GAT, and two GAT layers.Implementations used PyTorch and Deep Graph Library.Source: Baselines · p. 6

Evaluation Metrics

Prediction performance is measured with RMSE, MAE, and CPC. CPC measures agreement between predicted and target flows on a scale from no agreement to identical values.

Evaluation Metrics

  • The evaluation uses RMSE, MAE, and Common Part of Commuters (CPC) to measure prediction performance.The paper states that RMSE and MAE are common regression metrics and CPC is used in commuting-flow prediction.Source: Evaluation Metrics · p. 6
  • CPC ranges from 0 when predicted and target values have no agreement to 1 when they are identical.CPC measures the common part of agreements between predicted and target values.Source: Evaluation Metrics · p. 6

Performance Analysis

GMEL variants outperform the baselines, and full GMEL outperforms variants without multitask learning or separate supply-demand modeling. Gravity models perform worst; RF and GBRT generally do better than gravity models.

Performance Analysis

  • All GMEL variants outperform the baseline models, and full GMEL outperforms GMEL-noMul and GMEL-noSep.The comparison is summarized in Table 2.Source: Table 2 · p. 6, Performance Analysis · p. 7
  • Gravity models have the worst performance, while RF and GBRT generally perform better than gravity models.The authors suggest simple gravity-model assumptions may miss complex commuting patterns.Source: Performance Analysis · p. 7
  • Node2vec is slightly better than RF and comparable to GBRT, but its neighborhood information may not characterize urban supply and demand.The authors suggest that preserving network neighborhoods does not necessarily capture the city characteristics relevant to commuting flows.Source: Performance Analysis · p. 7

Residual Analysis

GMEL’s residuals are spatially smoother than GBRT’s, with the maps showing differences between predicted and ground-truth incoming flows by census tract.

Residual Analysis

  • GMEL’s incoming-flow residuals are spatially smoother than GBRT’s across census tracts.The maps show residual sums for flows to each destination; red indicates underestimation, blue overestimation, and light blue the best predictions.Source: Residual Analysis · p. 7, Figure 5 · p. 7

Parameter Sensitivity Analysis

The sensitivity analyses examine GAT depth, embedding size, multitask weights, salient urban indicators, and a cross-year case study of five census tracts.

Parameter Sensitivity Analysis

  • One GAT layer performs worse than two or more, whose performance changes little; two-hop neighborhoods span about 1.5 km on average in New York City.The authors equate this distance to approximately a 15-minute walk.Source: The effect of number of GAT layers · p. 7
  • Performance rises as embedding size increases from 32 and saturates at 128.The tested sizes were 32, 64, 128, and 256.Source: The effect of embedding size · p. 7
  • Main-task performance increases as subtask weights rise, reaching its best reported level when main-task and subtask weights are both 0.5.The subtasks are introduced to enhance the main task.Source: The effect of multitask weights · p. 8

Feature Sensitivity Analysis

  • Saliency maps identify flow-specific urban indicators: buildings per square meter for in-flows and residential floor-area ratio for out-flows.The maps use average output gradients with respect to urban indicators; larger absolute values indicate more prominent indicators.Source: Feature Sensitivity Analysis · p. 8, Figure 6 · p. 8

Case Study

  • For five tracts with the largest 2013–2015 indicator changes, a model trained on 2013 data was tested on 2015 data, with reported MAEs of 1.07 ± 1.25, 1.40 ± 2.12, 1.26 ± 2.08, 0.99 ± 1.73, and 0.86 ± 1.37.The case study describes using a model trained on one year to predict origins and destinations of new commuting flows given indicator changes.Source: Case Study · p. 8

Conclusion

The paper proposes geo-contextual embeddings for commuting-flow prediction from infrastructure and land-use information. On New York City data, the authors report improved accuracy and performance above all baseline methods.

Conclusion

  • The study predicts commuting flows using infrastructure and land-use information, a problem framed as relevant to urban planning and public policy.The conclusion identifies this as the paper’s problem setting.Source: Conclusion · p. 8
  • GMEL uses a graph-attention embedding-learning framework to encode geographic contextual information, then feeds the learned embeddings to a gradient boosting machine.The conclusion describes this as an end-to-end framework for learning geo-contextual embeddings.Source: Conclusion · p. 8
  • Experiments on real-world New York City data show that geographic contextual information improves prediction accuracy and GMEL outperforms all baseline methods, including the state of the art.Source: Conclusion · p. 8

Related papers

Loading 2005.01690v1…