基于多任务学习的快件送达时间预测方法  

Express Time Prediction Method Based on Multi-Task Learning

作　　者：王强[1] 林友芳[1,2] 万怀宇 WANG Qiang;LIN Youfang;WAN Huaiyu(School of Computer and Information Technology,Beijing Jiaotong University,Beijing 100044,China;Key Laboratory of Transport Industry of Big Data Application Technologies for Comprehensive Transport,Beijing 100044,China)

机构地区：[1]北京交通大学计算机与信息技术学院,北京100044 [2]综合交通运输大数据应用技术交通运输行业重点实验室,北京100044

出　　处：《计算机工程》2022年第4期314-320,共7页Computer Engineering

基　　金：国家自然科学基金(61603028)。

摘　　要：快件送货时间预测(即在任何时间预测包裹送达的到达时间)是物流领域中最重要的服务之一。准确地预测快件送达时间可以为用户提供更准时的服务,缓解客户的等待焦虑,提升用户体验,且有利于快递员的路径规划,从而提高派送效率。然而在快递派送场景下,多因素、动态性及多目的地等特征给快件投递准确预测送达时间带来巨大挑战。提出一种基于多任务学习的模型MTDTN,从快递员的大量历史时空轨迹中预测快件送达时间。MTDTN建模多种影响送达时间的外部因素,利用地理信息编码、卷积操作以及双向长短时记忆网络来捕获派送行为的时空关系,并运用多任务学习框架,引入顺序预测的辅助任务与送达时间预测的主任务,提高模型预测性能。在真实数据集上的实验结果表明,与基准方法中最优的DeepETA模型相比,该模型的平均绝对误差与平均绝对百分比误差分别降低了16.11%和12.88%,模型效果明显提升。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.

关 键 词：送达时间预测 时空轨迹 卷积神经网络 双向长短时记忆 多任务学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]