A Two-stage Learning-based method for Large-scale On-demand pickup and delivery services with soft time windows

•A novel two-stage learning-based method to solve the large-scale pick-up and delivery problems with soft time window.•Superior performance in numerical experiments comparing with tradition heuristic methods.•Computation time of high-quality solution within seconds even facing the problem of 500 customers.•Strong transferability of this method with much less training time and data.