Express Time Prediction Method Based on Multi-Task Learning

Abstract

Delivery time prediction(i.e., predicting package arrival time at any time) is important to logistics service providers.Accurate prediction of the delivery time provides customers with more prompt services and alleviates anxiety. It is also beneficial to the route planning by couriers for improved delivery efficiency.In real scenarios, however, accurate delivery time prediction is marred with multiple destinations, multiple factors, and dynamics challenges.In this paper, relying on the historical spatio-temporal trajectories of couriers, a Multi-Task model for Delivery Time prediction Network(MTDTN) is proposed to predict the package delivery time.MTDTN leverages external factors that may affect the delivery time and utilizes the geographic information encoder, convolution operation, and the Bidirectional Long Short-Term Memory(Bi-LSTM) to capture the spatio-temporal information in the trajectories.Moreover, multi-task learning is used to simultaneously predict both the delivery time and the delivery sequence.The model performance is enhanced by introducing the delivery sequence prediction as an auxiliary task.Experimental results on real data sets show that, compared with the optimal DeepETA model in the benchmark method, Mean Absolute Error(MAE) and Mean Absolute Percentage Error(MAPE) of this model are reduced by 16.11% and 12.88% respectively.