On-Road Vehicular Emission Measurement Data & Analysis leakage codes

We demonstrate this on a nine-vehicle Indian in-use fleet, showing that adding static vehicle attributes (kerb weight, engine displacement, odometer) raises random-fold R² to 0.60–0.86, but leave-one-vehicle-out R² drops to −0.47 to −1.64.

The mechanism: the three attributes jointly identify every vehicle uniquely, so the model functions as a vehicle-identity lookup rather than an emission model; random splitting leaks this identity across train/test boundaries. A decision-tree classifier recovers vehicle identity from these attributes with 100% accuracy. Pooling more training vehicles (2–8) does not close the gap.