With the increasing availability of big, transport-related datasets, detailed data-driven mobility analysis is becoming possible. Trips with their origins, destinations, and travel times are now collected in publicly available databases, allowing for detailed demand forecasting with methods exploiting big and accurate data. In this paper, we predict the demand pattern of New York City bikes with a low-dimensional approach utilizing three-level data clustering. We use historical demand data along with temperature and precipitation to first aggregate and then decompose data to obtain meaningful clusters. The core of this approach lies in the proposed clustering technique, which reduces the dimension of the problem and, differently from other machine learning techniques, requires limited assumptions on the model or its parameters. The proposed method allows, for the given temperature and precipitation method, to obtain expected vector of movement (mean number and direction of trips) for each zone. In this paper, we synthesize more than 17 million trips into daily and zonal vectors of movement, which combined with weather data allow forecasting of the trip demand. The method allows us to predict the demand with over 75% accuracy, as shown in series of experiments in which various settings and parameterizations are validated against 25% holdout data.

随着与运输相关的大型数据集可用性的提高，详细的数据驱动的移动性分析成为可能。现在，将带有起点，目的地和旅行时间的旅程收集在公开可用的数据库中，从而可以利用利用大而准确的数据的方法来进行详细的需求预测。在本文中，我们采用三级数据聚类的低维方法预测了纽约市自行车的需求模式。我们将历史需求数据与温度和降水一起使用，首先进行聚合，然后分解数据以获得有意义的聚类。该方法的核心在于提出的聚类技术，该技术减少了问题的范围，并且与其他机器学习技术不同，它要求对模型或其参数进行有限的假设。对于给定的温度和降水量方法，所提出的方法可以获取每个区域的预期运动矢量（均值和行程方向）。在本文中，我们将超过1700万次旅行合成为每日和区域运动矢量，结合天气数据可以预测旅行需求。该方法使我们能够以超过75％的准确度预测需求，如一系列实验所示，其中针对25％保持数据验证了各种设置和参数设置。