Crowd-shipping (CS) is an innovative logistics service that occasional and professional couriers sign up for via an online platform to deliver packages upon requests by senders. Currently the demand and supply of CS is not yet well understood, largely due to the limited real-world data. This study aims to first fill this gap by analyzing the real-world CS data from the city of Atlanta, GA between April 2015 and August 2018. We first present an overview of the real-world CS data in three aspects: (1) the CS pricing scheme; (2) the CS spatial and temporal delivery patterns; and (3) comparison of preferences between the senders’ requests and the couriers’ bids. The analysis finds that the CS service has a clear price advantage over FedEx in the same-day and express service, as well as in the large, extra large, and huge size package delivery. The data analysis also reveals considerable discrepancies between senders’ and couriers’ preferences. We then compare two classes of the state-of-the-art Deep Learning (DL) methods in their ability to predict short-term CS delivery trip production. One class captures only the temporal features, namely the Long Short-term Memory Neural Network (LSTM), the Bidirectional Long Short-term Memory Neural Network (BDLSTM), and the Gated Recurrent Unit (GRU). The other class considers both spatial and temporal features, namely Convolutional Neural Network (CNN), CNN-LSTM, and ConvLSTM. The results show that ConvLSTM has overall the best predictive performance among the six DL methods considered, proving the importance of capturing both the spatial and temporal features of the delivery trip production data, as well as the convolutional nature of the spatial and temporal features in the data.

群众运输（CS）是一项创新的物流服务，偶尔的和专业的快递员通过在线平台进行注册，以根据发件人的要求交付包裹。目前，对CS的需求和供应还没有很好的了解，这在很大程度上是由于实际数据有限。这项研究旨在通过分析2015年4月至2018年8月间佐治亚州亚特兰大市的真实CS数据来填补这一空白。我们首先从以下三个方面对真实CS数据进行概述： CS定价方案；（2）CS的空间和时间传递模式；（3）比较发件人的请求和快递员的出价之间的偏好。分析发现，CS服务在当日和快递服务以及大件，超大件和大件包裹递送中具有明显优于联邦快递的价格优势。数据分析还揭示了发件人和快递员的偏好之间的巨大差异。然后，我们比较两类最先进的深度学习（DL）方法在预测短期CS交付行程产生中的能力。一类仅捕获时间特征，即长短期记忆神经网络（LSTM），双向长短期记忆神经网络（BDLSTM）和门控循环单元（GRU）。另一类同时考虑空间和时间特征，即卷积神经网络（CNN）分别是长期短期记忆神经网络（LSTM），双向长期短期记忆神经网络（BDLSTM）和门控循环单元（GRU）。另一类同时考虑空间和时间特征，即卷积神经网络（CNN）分别是长期短期记忆神经网络（LSTM），双向长期短期记忆神经网络（BDLSTM）和门控循环单元（GRU）。另一类同时考虑空间和时间特征，即卷积神经网络（CNN）， CNN-LSTM和ConvLSTM。结果表明，ConvLSTM在所考虑的六种DL方法中总体上具有最佳的预测性能，证明了捕获交付行程生产数据的时空特征以及该时空特征的卷积性质的重要性。数据。