Due to its various social and environmental benefits, bike sharing has been gaining popularity worldwide and, in response, many cities have gradually expanded their bike sharing systems (BSSs). For a growing station-based BSS, it is essential to plan new stations based on existing ones, which requires predicting not only the overall trip intensity at each station but also its temporal distribution, an issue underexplored in the literature. To this end, this study investigates the problem of time-dependent trip generation for BSS planning (TTGP), which aims to forecast the number of trips generated by new stations at different time periods. This task, however, is challenging due to the lack of historical data for newly planned stations and complex spatiotemporal dependencies in bike sharing demand. To address these challenges, we propose a multi-task memory-augmented graph neural network for TTGP by leveraging its surrounding urban contexts and the historical demand features of nearby existing stations. Specifically, a feature extractor is developed, consisting of a graph neural network and a memory network to encode urban context and historical demand features, respectively, and a gate network to learn the reliability of different features. Furthermore, a multi-task demand predictor is designed to predict the daily trip intensity and its hourly distribution as two distinct tasks. Finally, extensive experiments on real-world data from New York City demonstrate the superior performance of our method compared with existing baselines.

中文翻译：

---

自行车共享规划的时间相关行程生成：多任务记忆增强图神经网络

由于其各种社会和环境效益，共享单车在全球范围内越来越受欢迎，为此，许多城市逐渐扩大了自行车共享系统（BSS）。对于不断发展的基于站点的BSS，必须在现有站点的基础上规划新站点，这不仅需要预测每个站点的总体出行强度，还需要预测其时间分布，这是文献中尚未探讨的问题。为此，本研究研究了BSS规划（TTGP）的​​时间相关出行生成问题，旨在预测不同时间段新车站生成的出行数量。然而，由于缺乏新规划站点的历史数据以及自行车共享需求复杂的时空依赖性，这项任务具有挑战性。为了应对这些挑战，我们利用周围的城市环境和附近现有车站的历史需求特征，为 TTGP 提出了一种多任务记忆增强图神经网络。具体来说，开发了一种特征提取器，由图神经网络和记忆网络组成，分别用于编码城市背景和历史需求特征，以及门网络来学习不同特征的可靠性。此外，多任务需求预测器旨在将每日出行强度及其每小时分布预测为两个不同的任务。最后，对纽约市真实世界数据的广泛实验证明了我们的方法与现有基线相比具有优越的性能。