Explainable machine learning for trip-level driving behaviour profiling from smartphone telematics in road freight operations

Low-cost smartphone telematics can support transparent trip-level profiling in road freight operations, but telemetry-derived labels require careful interpretation. We analyse an extract of 52 freight trips by 12 drivers (11,434.91 km) associated with a platform reporting 753,021 source location records. Distance-normalised acceleration, braking and route-alert indicators were used to construct four rule-based operational tiers (Excellent, Good, Moderate and Poor); K-means was used for exploratory description only. Decision tree (DT), random forest (RF) and k-nearest neighbours were evaluated using a driver-grouped hold-out (37 training and 15 test trips) and leave-one-driver-out cross-validation. In the hold-out, RF achieved accuracies of 0.867 for driving behaviour and 0.933 for the eco-driving proxy. Removing the score-generating indicators and their algebraically related features reduced RF accuracy to 0.467 and 0.533, respectively. The corresponding leave-one-driver-out RF accuracies were 0.942 and 0.962 using the full feature set, decreasing to 0.385 and 0.365, respectively, under strict feature ablation. Class-specific errors and local surrogate explanations further characterised model behaviour. The results support reproducible assignment of an operational telemetry taxonomy, not independently validated predictions of driver safety, energy consumption or emissions. The small driver cohort and missing vehicle, traffic and payload information limit generalisation and causal interpretation.