As mentioned on the competition page, we can not use the arrival airports data while predicting pushback time. Does it mean that we can not even build a model that has seen all the train data from all the airports? shall we have 10 different models for each of the airports or a single model is still fine?
The decision about whether it is better to train one model per airport or a single model for all airports is up to you. The constraint stated here is that when making a prediction for a flight at an airport, you may only use features that are included in that airport’s data directory. (This constraint is in fact strictly enforced in the prescreened code execution runtime.) The model itself may have been trained on data from all airports, however.