Attention Enhanced Package Pick-Up Time Prediction via Heterogeneous Behavior Modeling

The logistics industry has developed rapidly with the popularity of online-to-offline businesses in recent years. First-mile package pick-up is one of the most critical and expensive parts of the whole logistics service chain, which is finished by couriers in practice. Accurate prediction of package pick-up time at the customers’ addresses is essential to improve customers’ experience and increase platforms’ profits. For some logistics service providers, couriers conduct heterogeneous tasks (i.e., first-mile pick-up and last-mile delivery) simultaneously in a certain area to improve efficiency. However, existing works neglect the impact of the package delivery process, which produces inaccurate prediction results due to the coupling of the pick-up and delivery process. Considering the delivery process in pick-up time prediction introduces two additional challenges: (i) Limited pickup requests. In practice, couriers have a limited number of package delivery tasks in a delivery trip, which hinders the direct application of existing deep learning models for the prediction. (ii) Dynamic package pickup requests. Package pick-up requests are generated dynamically, which affects the courier’s route. In this paper, we propose HTAPT, a heterogeneous tasks aware package pick-up time prediction framework, which consists of two modules: (i) Pre-trained stay time prediction module to learn the embedding of the courier’s stay time. (ii) Attention enhanced pick-up time and route prediction module to predict the delivery route and pick-up arriving time of the courier under the pick-up influence. The evaluation results with real-world order data from JD Logistics, which is one of the largest logistics companies in China show HTAPT improves the prediction accuracy by up to 10% compared with the state-of-the-art methods.