Travel Time Prediction from Sparse Open Data

AI summary

Overview

Research area: Transport geography / GIScience — travel time prediction and urban accessibility analysis.

Technical level: Intermediate. The methods (random forest regression, cross-validation, SHAP) are standard machine learning tools, but the paper is written for planning practitioners rather than ML specialists and does not require deep learning background.

Scope: The paper proposes and validates a low-cost, open-data random forest model that predicts minimally-congested driving travel times across a metropolitan area, offering a middle ground between crude naïve routing and expensive, data-hungry state-of-the-art models.

What This Paper Is About

Travel time is essential for accessibility studies and transport planning, but accurate driving times are hard to obtain: sophisticated models need massive proprietary datasets or deep-learning infrastructure, while simple "naïve" methods (minimizing distance or speed-limit traversal time) systematically under-predict real trips. The authors ask whether free open data — street networks, speed limits, turn directions, and traffic control features — plus a small one-time sample of Google travel times can train a lightweight model that approaches the accuracy of much more resource-intensive approaches. They test this idea on the Los Angeles urban area.

Key Contributions

1. A minimal-resource prediction framework. A complete open-source pipeline that predicts point-to-point driving travel times using only OpenStreetMap data, free open-source Python tooling, and a small one-off sample of Google Routes API queries for training — no bespoke GPS data, deep learning stack, or proprietary licenses.

2. A quantified "middle ground" benchmark. The paper formalizes three criteria for a practical middle-ground method (free data/software, ease of use, better accuracy than naïve methods with point-to-point resolution) and empirically demonstrates a model that satisfies all three, unlike either the naive planning literature or the state-of-the-art engineering literature.

3. Demonstration that sparse open data suffices. Traffic control tagging in OpenStreetMap is far from complete, yet the model still reaches high accuracy — showing that partial open data on turns and traffic controls carries substantial predictive signal.

4. A robustness and interpretability analysis. The workflow is re-run on a later (2025) OpenStreetMap snapshot with nearly double the traffic-control coverage and yields equivalent performance, and SHAP values quantify how much each feature type contributes to predictions.

Main Findings

• Random forest achieves state-of-the-art-level accuracy with minimal inputs. Out-of-sample MAPE of 8.42%, MAE of 75.3 seconds, MSE of 12,182 s², and R² of 0.93 — comparable to the roughly 3–17% MAPE range reported by deep-learning and graph neural network studies that consume millions or billions of records.

• Naïve methods are badly biased. The baseline (shortest path minimizing speed-limit traversal time) had a MAPE of 21.15%, an MAE of 183.7 seconds, and systematically under-predicted Google travel times by 182.9 seconds (about 3 minutes) on average, with an average pairwise ratio of 0.79 and R² of 0.74.

• The chosen model has essentially no systematic bias. The random forest over-predicted by only 0.34 seconds on average (p = 0.78, statistically indistinguishable from zero), with an APR of 1.01. Gradient boosting and AdaBoost had slightly better MAPE but significantly under-predicted (δ = −19.3 s and −9.8 s), which is why the random forest was selected.

• Error reduction factors. Relative to the naïve baseline, the model cut MAE by a factor of 2.4 and MSE by a factor of 4.0.

• Results are robust to data updates. Re-running on 2025 OpenStreetMap data (with traffic control elements rising from 63,359 to 121,442) produced a random forest MAPE of 8.12%, MAE of 73.2 s, δ of 0.15 s (p = 0.90), and R² of 0.94 — and the same optimal hyperparameters emerged from the grid search.

• No overfitting. Five-fold cross-validation MAEs were tightly clustered (e.g., 75.4, 74.5, 73.7, 73.2, 74.7 for the 2023 model).

• Feature importance follows transport theory. SHAP analysis showed naïve travel time dominates (mean absolute contribution 328.1 s), followed by pedestrian crossings (20.2 s), left turns (5.1 s), traffic signals (4.9 s), right turns (3.7 s), and stop signs (2.8 s). Give-way signs and mini roundabouts contributed negligibly. More crossings, turns, and signals all increased predicted travel time.

• Tuned configuration is simple. The final random forest used 400 trees, a maximum depth of 10, all features considered at each split, and a minimum of two samples per split.

Methodology in Plain English

The authors build a street network graph for the Los Angeles urban area from OpenStreetMap, containing about 783,000 nodes and 127,000 intersections, and tag each node with whatever traffic controls are recorded (stop signs, signals, crossings, give-way signs, mini roundabouts).

To create a realistic sample of trips, they draw 5 million random origin-destination node pairs and filter them using Uber Movement data, keeping only pairs where real trips actually occurred. They restrict to trips in the 3:00 a.m. hour to approximate uncongested conditions, leaving about 41,000 OD pairs.

For each pair they compute a "naïve" travel time using Dijkstra's algorithm minimizing speed-limit traversal time, and count how many traffic controls and how many turns of each type (left, slight left, right, slight right, u-turn) that route passes through. They then query the Google Routes API for each pair just before 3:00 a.m. to get a traffic-aware travel time, treating this as the ground truth.

Finally, they train several tree-based models (decision trees, random forest, gradient boosting, AdaBoost) to predict the Google travel time from 11 predictors: the naïve time plus the five traffic-control counts plus the five turn counts. Hyperparameters are tuned with a randomized grid search and 5-fold cross-validation, model selection balances accuracy against systematic bias, and SHAP values explain feature contributions.

Why This Matters

Impact on research: The paper exposes a blind spot in the planning literature — the widespread use of naïve routing that under-predicts real travel times by minutes — and supplies a validated, reproducible alternative. Because it avoids proprietary or bespoke data, it lowers the barrier to entry for under-resourced researchers and makes accessibility findings more comparable and credible across studies.

Real-world applications:

• Scenario planning: Estimating how adding a street, changing a traffic signal, or redesigning an intersection would affect travel times, which closed black-box APIs cannot support.
• Equity and access analysis: Computing travel times to jobs, schools, or healthcare across an entire metropolitan area without per-query API costs.
• Congestion baselines: Establishing a congestion-free reference against which the accessibility costs of peak-hour traffic can be measured.
• Community advocacy: Giving neighborhood groups and small municipalities the tooling to contest or support infrastructure proposals with quantitative travel-time evidence.

Industry relevance: Commercial routing vendors and metropolitan planning organizations may find the approach useful for filling gaps where detailed traffic data are unavailable. The finding that sparse, imperfect open data still yields high accuracy is encouraging for crowdsourced geodata platforms and for firms building low-cost routing products for smaller markets or developing regions.

Future Directions

• Generalizability testing. The model was trained on Los Angeles; it is unknown how well it transfers to cities with sparser, lower-quality, or differently structured OpenStreetMap data, and whether retraining on local data is sufficient (as Breiman's principle suggests).

• Broader validation data. Google travel times are a black box derived from user GPS traces; future work should train and validate against independent empirical travel time datasets, including direct vehicle observations.

• Extension to congested conditions. The same framework could be applied to other times of day if congestion data were available, though the availability of such data is precisely the constraint the paper is designed to work around.

• Wider model scope. The authors suggest extending the approach to multimodal travel times and to more varied trip purposes, and exploring how the framework behaves as open traffic-control tagging improves over time.

Target Audience

Urban planning researchers and practitioners, transport geographers, and GIScience analysts who need defensible travel time estimates without access to proprietary data or high-performance computing. It is also valuable for students learning applied machine learning in a spatial context, and for policy analysts at smaller agencies or advocacy organizations who currently rely on distance-based or simple network-based accessibility measures. Computer scientists and engineers working at the deep-learning frontier of travel time prediction are less likely to find novel methodology here, but may appreciate the resource-accuracy trade-off framing.

Authors’ abstract

Travel time prediction is central to transport geography and planning's accessibility analyses, sustainable transportation infrastructure provision, and active transportation interventions. However, calculating accurate travel times, especially for driving, requires either extensive technical capacity and bespoke data, or resources like the Google Maps API that quickly become prohibitively expensive to analyze thousands or millions of trips necessary for metropolitan-scale analyses. Such obstacles particularly challenge less-resourced researchers, practitioners, and community advocates. This article argues that a middle-ground is needed to provide reasonably accurate travel time predictions without extensive data or computing requirements. It introduces a free, open-source minimally-congested driving time prediction model with minimal cost, data, and computational requirements. It trains and tests this model using the Los Angeles, California urban area as a case study by calculating naive travel times from open data then developing a random forest model to predict travel times as a function of those naive times plus open data on turns and traffic controls. Validation shows that this interpretable machine learning method offers a superior middle-ground technique that balances reasonable accuracy with minimal resource requirements.