Predicting when delivery drivers go off-route

Route planners give drivers an optimised stop sequence, but drivers often follow their own order. My thesis internship at Urbantz asked whether those deviations can be predicted, and whether planned routes can be adjusted so drivers actually stick to them.

I engineered 75+ features from GPS traces, timestamps and operational data, and built a three-phase XGBoost framework that predicts deviations at route and edge level. To prevent data leakage I used time-based train/test splits, and I validated the models with calibration analysis and rank correlations (Kendall τ, Spearman ρ). A greedy local-search algorithm then used the learned driver preferences to generate new stop sequences.

The models reached an AUC of 0.73–0.87 at route level and 0.66–0.73 at edge level. The optimised sequences cut predicted deviation risk on 96.5% of candidate routes. Thesis grade: 8.5/10.