PLSRP: prompt learning for send-receive path prediction

Qi Wei, Yi-Hong Lu, Dan Du, Huan-Tao Cai, Pei-Yuan Lai, Chang-Dong Wang

Abstract: With the rapid growth of the courier industry driven by societal development, efficiently and accurately predicting the send–receive path has become a critical issue. Existing methods suffer from several limitations, including discrepancies between pretraining and downstream data, and inconsistencies between training and testing targets. To address these challenges, this paper introduces a pioneering approach that applies prompt learning to send–receive path prediction. The proposed method employs a “pretraining-prompt-finetuning" paradigm, where a model is pretrained on a large-scale dataset and then finetuned using prompt vectors to adapt to downstream tasks. This novel strategy effectively bridges the gap between pretraining and finetuning data, ensuring better model generalization with minimal additional cost. Furthermore, we incorporate an actor-critic reinforcement learning framework, where the actor network generates paths and the critic network evaluates them. This framework optimizes the model based on rewards calculated from non-differentiable test criteria, effectively addressing the inconsistency between training and testing objectives. This approach is better suited to adapt to various delivery scenarios, enhancing prediction accuracy and efficiency. Experiments conducted on two real-world datasets demonstrate the superiority of the proposed method.