Taxi passenger demand prediction is of great significance to perceive citywide human mobility and make a lot of urban sensing applications more convenient. There are two major challenges to develop accurate predictive models, i.e., the complexity of the spatial-temporal dependencies as well as the dynamicity caused by some unpredictable dependencies. Although existing work uses various methods such as time series analysis, machine learning, and deep learning, most of them ignore two facts: the uncertainty of taxi demands and the impact of the parallel car-hailing markets (e.g., Uber demands) on taxi demands. In this paper, in order to deal with these two facts systematically, we design a unified framework that can use multi-source data to improve prediction accuracy. Specifically, we analyze the correlations between taxi and Uber demands and design two deep models, each of which containing a specific feature fusion method. The first model adaptively aggregates features of each grid according to the correlations. To realize the feature fusion among adjacent grids, the other method contains an additional local convolution. Besides, we also study the impact of Uber demand trends on taxi demands and aggregate the impact into the second model to improve prediction accuracy. We evaluate our models based on both taxi and Uber datasets collected from New York City, USA. Results show that our models achieve superior performance compared to the state of the art.

出租车乘客需求预测对于感知整个城市的人员流动和使许多城市传感应用更加便捷具有重要意义。开发准确的预测模型存在两个主要挑战，即时空依赖性的复杂性以及某些不可预测的依赖性导致的动态性。尽管现有工作使用时间序列分析，机器学习和深度学习等各种方法，但大多数方法都忽略了两个事实：出租车需求的不确定性以及并行的叫车市场（例如，Uber需求）对出租车需求的影响。在本文中，为了系统地处理这两个事实，我们设计了一个统一的框架，该框架可以使用多源数据来提高预测准确性。特别，我们分析了出租车和Uber需求之间的相关性，并设计了两个深度模型，每个模型都包含一个特定的特征融合方法。第一模型根据相关性自适应地聚合每个网格的特征。为了实现相邻网格之间的特征融合，另一种方法包含附加的局部卷积。此外，我们还研究了Uber需求趋势对出租车需求的影响，并将影响汇总到第二个模型中以提高预测准确性。我们基于从美国纽约市收集的出租车和Uber数据集评估模型。结果表明，与现有技术相比，我们的模型具有出色的性能。为了实现相邻网格之间的特征融合，另一种方法包含附加的局部卷积。此外，我们还研究了Uber需求趋势对出租车需求的影响，并将影响汇总到第二个模型中以提高预测准确性。我们基于从美国纽约市收集的出租车和Uber数据集评估模型。结果表明，与现有技术相比，我们的模型具有出色的性能。为了实现相邻网格之间的特征融合，另一种方法包含附加的局部卷积。此外，我们还研究了Uber需求趋势对出租车需求的影响，并将影响汇总到第二个模型中以提高预测准确性。我们基于从美国纽约市收集的出租车和Uber数据集评估模型。结果表明，与现有技术相比，我们的模型具有出色的性能。