It has been gradually recognized that mobile phones can be used as a practical and promising way to identify individual travel trajectories. Researchers have developed various approaches to detecting human mobility and trip characteristics including trip origin–destination, travel modes, trip purposes based on mobile phone data. Among these researches, trip purpose detection has drawn less attention from researchers. This paper presents our work to investigate a set of machine learning approaches to identifying the trip purposes for tourists based on mobile signaling data combined with sampling surveys and point of interest (POI) data. Five machine learning algorithms, including support vector machine, decision tree, random forest, artificial neural network, and deep stacked auto-encoded (DSAE), have been employed to infer trip purposes under multiple scenarios. Four scenarios have been designed by considering the POI information around trip end [a 500 m buffer or Thiessen polygon (the coverage of the base station theoretically)] and training dataset selection (equal probabilities selection or equal proportion selection). The accuracy of trip purpose classification with machine learning algorithms has compared under different scenarios. The highest accuracy of 93.47% for the test dataset is achieved based on DSAE model under the scenario of a trip end 500 m buffer and equal probabilities selection. The experimental results indicate that the methodology developed with machine learning algorithms based on mobile signaling data combined with sample travel survey is expected as an alternative way to traditional travel surveys for trip purposes.

人们逐渐认识到，手机可以作为一种实用且有前景的识别个人旅行轨迹的方式。研究人员开发了各种方法来检测人类移动性和旅行特征，包括基于手机数据的旅行起点-目的地、旅行模式、旅行目的。在这些研究中，旅行目的检测较少受到研究人员的关注。本文介绍了我们的工作，以研究一组机器学习方法，以根据移动信号数据结合抽样调查和兴趣点 (POI) 数据来确定游客的旅行目的。五种机器学习算法，包括支持向量机、决策树、随机森林、人工神经网络和深度堆叠自动编码（DSAE），已被用于推断多种情景下的旅行目的。通过考虑行程终点附近的POI信息[500 m缓冲区或泰森多边形（理论上基站的覆盖范围）]和训练数据集选择（等概率选择或等比例选择），设计了四种场景。使用机器学习算法对旅行目的分类的准确性在不同场景下进行了比较。在行程终点500 m缓冲区和等概率选择的场景下，基于DSAE模型的测试数据集的最高准确率为93.47%。