Project Detail

French LCV Traffic Modeling (Paris)

Built a two-part ML pipeline for Paris light commercial vehicle analysis: (1) stop-location prediction and (2) stop-duration prediction.

Problem

Urban freight analysis required more than origin-destination aggregates. The core need was to estimate where LCVs are likely to stop and how long they remain there, using noisy, large-scale mobility traces enriched with land-use and economic context.

Approach

1) Stop location prediction - Prepared LCV trip start/end flows and mapped them to IRIS zones. - Built POI candidate sets using KNN + zone-based expansion. - Engineered temporal/spatial/context features (hour/weekday, road type/direction/modes, POI category context, zone-level POI and job frequencies). - Trained an XGBoost model with hard/soft labeling and weighted training. 2) Stop duration prediction - Built a duration feature table from large stop-event data (over 1.3M rows). - Added nearest-POI context, IRIS logistics features, and employment/POI zone features. - Created train/valid/test splits and trained an XGBoost regressor for stop duration. Data foundation across both tasks: LCV flow/stop data, POI datasets, IRIS boundaries, road-network attributes, and employment/business activity aggregates.

Results

Delivered an end-to-end modeling workflow that produced stop-level location probabilities and duration estimates, plus reusable feature pipelines and model artifacts for repeatable urban freight and mobility analysis.