Estimating Pedestrian Volumes from GIS-Derived Built-Environment Features: A Machine Learning Framework
Bahareh Golchin, Banafsheh Rekabdar, Sirisha Kothuri, Joseph Broach
cs.LG
Quick verdict
Claims tuned HistGB-Poisson + Lasso cuts CV RMSE 12% vs NB GLM on 101 Portland counts; careful pipeline, but 27% in conclusion mixes holdout and CV protocols.
Summary
The paper predicts 2-hour PM peak pedestrian volumes at 101 intersections in Portland, Oregon from GIS-derived features, targeting the small-sample regime (81 training sites versus 92 encoded predictors) typical of direct-demand modeling. Counts were manually extracted from 192 City of Portland PDFs and cluster downtown and in spring and summer, so coverage is not random. Starting from a Negative Binomial GLM baseline, the authors benchmark seven models, four feature selectors, four cross-validation strategies, and five feature-count settings (roughly 70,000 fits), selecting one configuration by combined rank over RMSE, MAPE, and SMAPE. The winner — histogram gradient boosting with Poisson loss, L1-Lasso selection of 20 features, tuned under repeated 3×10 CV — reaches CV RMSE 78.7 versus the GLM's 89.8 (12% lower) and holdout RMSE 87.9 versus 108.0 (19% lower). The negative results are useful: without feature selection the GLM over-predicts 254-fold; log-transforming the target damages Poisson-loss boosting; and stacking loses to a tuned single model. Main limitations: the conclusion's 27% figure mixes a holdout baseline with a cross-validated winner; 526-configuration selection carries no confidence intervals or nested-validation correction; and the top predictor is Strava crowdsourced walking activity — an activity proxy rather than a built-environment feature, straining the open-GIS framing.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A tuned histogram gradient boosting model with Poisson loss plus L1-Lasso selection of 20 features reduces cross-validated RMSE from 89.8 to 78.7 (−12%) and holdout RMSE from 108.0 to 87.9 (−19%) versus a Negative Binomial GLM on 101 Portland intersections.
- Feature selection is mandatory in this regime: all 92 encoded features against 81 training samples drive the GLM to a training pseudo-R² of 1.000 and 254× over-prediction on test sites.
- Log-transforming the target degrades Poisson-loss gradient boosting (73.7→106.8) because of the double log-link, and Ridge-meta stacking (best 76.1) fails to beat the tuned single model.
- The selected predictors (Strava activity, CBD distance, signalization, retail floor area, zero-car households, sidewalk length) align with established pedestrian planning theory, supporting conceptual validity.
- The conclusion claims a 27% RMSE reduction over the published baseline — a figure that mixes a holdout baseline with a cross-validated winner.
Strengths
Validation hygiene is genuinely careful for n=101
Selectors and all preprocessing are fit on training folds only (§3.2–3.3); repeated 3×10 CV was chosen because it gave the smallest RMSE SD across the 32 design cells; a quintile-stratified CV rerun (RMSE 81.05, ranking unchanged) probes heavy-tail sensitivity.
Anti-cherry-picking model selection
The winner is chosen by average rank over RMSE, MAPE, and SMAPE across all four CV strategies, retaining only top-decile candidates — exactly one of 77 configurations survives, and Table 1 honestly labels which numbers are holdout versus CV.
Practitioner-useful negative results
Reproducibility specification
Conceptual validity check against planning theory
Weaknesses
Cross-protocol headline inflation
Finding 1 is unsupported under CV
No uncertainty quantification or nested validation
Strava covariate dominance strains the framing
Internal number drift and narrow external validity
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is a 4-page SIGSPATIAL '26 short paper predicting 2-hour PM peak pedestrian volumes at 101 intersections in Portland, Oregon, from 79 GIS-derived candidate features (92 after one-hot encoding) drawn from OpenStreetMap, ACS, LEHD, and Strava Metro. The stated contribution is an end-to-end pipeline for the small-sample regime (81 training sites): a Negative Binomial GLM baseline (published R-code, 10 features) plus feature selection, count-aware gradient boosting, repeated CV, and a combined-rank selection rule, evaluated over roughly 70,000 fits across 526 configurations. Central claims: (1) the winner — HistGB with Poisson loss, L1-Lasso selection of 20 features, tuned hyperparameters — cuts cross-validated RMSE from 89.8 to 78.7 (−12%) and holdout RMSE from 108.0 to 87.9 (−19%); (2) feature selection is mandatory; (3) log-target transformation and stacking do not help; (4) the conclusion escalates the gain to 27% — a figure this review contests.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
How much better is the machine learning model than the standard GLM?
Under identical repeated 3×10 cross-validation, RMSE improves from 89.8 to 78.7 (12% lower) and MAPE from 66.1% to 51.8%; on the 20-site holdout, RMSE improves from 108.0 to 87.9 (19% lower). The conclusion's 27% figure compares the holdout baseline with the cross-validated winner and mixes the two protocols.
What data does the pedestrian volume model use?
Two-hour PM peak (4–6 PM) manual counts at 101 Portland, Oregon intersections from 192 City of Portland PDFs, plus 79 GIS-derived predictors (92 after one-hot encoding) from OpenStreetMap, US Census/ACS, LEHD, and Strava Metro activity, aggregated at 100-foot intersection buffers and at 1/8, 1/4, and 1/2 mile.
Can I apply this pipeline to my own city?
The code is on GitHub, but expect recalibration: the paper's 101 sites are non-random (downtown and corridor-heavy), counts are spring/summer only, and the top predictor is Strava crowdsourced walking activity with known demographic bias — validate against local counts before trusting network-wide estimates.
Should I log-transform counts before Poisson gradient boosting?
This paper says no: log-transforming the target hurt Poisson-loss boosting badly (RMSE 73.7 to 106.8) because the Poisson loss already applies an internal log link, so pre-transforming creates a double log-link that distorts the conditional mean.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–3 are raster images; only captions and the inline numeric labels reproduced in the text (e.g., Figure 3's ΔRMSE values) were extractable, so the Figure 1 map's spatial pattern cannot be independently assessed.
- The GitHub repository was not inspected; 'code released' reflects the paper's own statement, and whether the 101×92 feature table ships with it is unknown.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.