A courier’s choice for delivery gigs in a crowdshipping service is not well understood in the literature. Thus the objective of this study is to empirically investigate the crowdshipping (CS) couriers’ bidding preferences for delivery gigs, and how the gig features impact the gig delivery status of a real-world CS service in the United States. The delivery records were made available between 2015 and 2018. A descriptive analysis reveals that there exist significant preference discrepancies between the senders and the couriers in terms of package size, delivery time window, delivery distance, and delivery fee. Therefore, four features to capture the above discrepancy are specifically created from the data in predicting the bidding level and the delivery status. The bidding level which is measured by the number of bids received per gig is classified into low, medium, and high bidding levels to reflect the couriers’ preferences for the delivery gigs. The delivery status, labeled as delivered or undelivered, is affected by the couriers’ eventual choice of the delivery gigs. Five popular machine learning (ML) methods, namely Random Forest Decision Tree, Artificial Neural Network, eXtreme Gradient Boosting (XGBoost), Support Vector Machine, and Bayesian Network are applied to the predictions. Among them, the XGBoost is found to perform the best. Furthermore, the Shapley Additive exPlanations (SHAP) values are introduced to explain and visualize how each feature influences the dependent variable (prediction target). The SHAP values provide an effective visualization and interpretability of the feature impact values and importance rankings, much like the coefficients of the traditional econometric based logit model. The paper further demonstrates that the ML models and the logit models produce consistent feature influences. Overall, the couriers are generally interested in the delivery gigs of extra-large and huge package sizes, medium to long delivery distance, insured packages, and flexible delivery time window. Discrepancy related features significantly influence couriers’ bidding behavior as expected. The study also reveals that gigs that receive a high number of bids do not translate into their eventual successful deliveries. Finally, policy and practical implications for improving the CS service particularly through pricing strategies are discussed.

文献中并没有很好地理解快递员在众包服务中选择送货方式。因此，本研究的目的是实证调查众包 (CS) 快递员对送货服务的出价偏好，以及送货功能如何影响美国现实世界 CS 服务的送货状态。公开了2015年至2018年的投递记录。描述性分析显示，寄件人和快递员在包裹尺寸、投递时限、投递距离、投递费用等方面存在显着的偏好差异。因此，专门根据数据创建了四个特征来捕获上述差异，以预测投标水平和交付状态。竞价级别以每场比赛收到的竞价数量来衡量，分为低、中、高竞价级别，以反映快递员对送货场次的偏好。投递状态（标记为已投递或未投递）受到快递员最终选择的投递地点的影响。五种流行的机器学习（ML）方法，即随机森林决策树、人工神经网络、极限梯度提升（XGBoost）、支持向量机和贝叶斯网络被应用于预测。其中，XGBoost 表现最好。此外，还引入了 Shapley 加性解释 (SHAP) 值来解释和可视化每个特征如何影响因变量（预测目标）。 SHAP 值提供了特征影响值和重要性排名的有效可视化和可解释性，非常类似于基于传统计量经济学的 Logit 模型的系数。 该论文进一步证明了 ML 模型和 Logit 模型产生了一致的特征影响。总体来看，快递员普遍对特大件、巨型包裹、中远距离、有保险的包裹、灵活的时效窗口感兴趣。差异相关特征显着影响快递员的预期出价行为。该研究还表明，收到大量投标的演出并不能转化为最终的成功交付。最后，讨论了改善客户服务服务的政策和实际影响，特别是通过定价策略。