面向地铁客流大数据的统计机器学习关键技术研究

Statistical machine learning plays a significant role in analyzing and modeling metro passenger flow big data, especially in the process of mining hidden patterns and useful information from data. Featured by multi-modal heterogeneity, spatial and temporal distribution diversity and personalized information, metro passenger flow data brings challenges to various machine learning tasks, including passenger flow distribution analysis, travelling patterns recognition, and passenger flow forecasting, etc. The proposed research focuses on addressing several key difficulties in metro passenger flow big data modeling and analysis via statistical machine learning. The proposed research is in two major areas: considering the complex network structure of metro system, heterogeneity of passenger traveling patterns, and data streams from multiple sources, we will develop effective clustering methods for high-dimensional sparse data, aiming to provide solutions for station clustering and passenger travel behavior analysis; besides, we will conduct research on feature extraction strategy via data integration and develop predictive models for passenger flow forecasting by structuring unstructured data, which would enable efficient passenger flow monitoring as well as influencing factors identification. Our research outcomes will create academic values in delivering dynamic forecasting models, high-dimensional clustering algorithms, and informed decision support approaches for complex metro networks under the big data environment. On the other hand, our research will contribute to information management and transportation planning in metro system and data driven analytics for smart city initiatives.

统计机器学习技术在地铁客流大数据的分析与处理，特别是在挖掘其中的隐含模式和有价值信息过程中起着重要作用。然而，地铁客流大数据的多模态异构、时空分布差异性显著和信息个人化等特征为传统统计机器学习技术在挖掘客流分布及乘客出行模式、预测客流量等问题上带来了新的挑战。为此，本项目拟面向地铁客流大数据开展统计机器学习的若干关键技术问题研究，包括：考虑地铁复杂网络结构和乘客出行的规律多样性以及时空分布异质性，对多源多模态数据流的特征学习；研究针对高维数据的高效聚类算法，用以解决站点类型识别和乘客出行聚类模式分析；在此基础上，研究基于数据融合的特征提取策略，对非结构化数据进行分析和建模，开发客流量预测模型并进行影响因素甄别。本项目的研究将在理论上解决大数据环境下复杂网络中的动态预测问题和集群识别问题，具有重要的学术价值，在地铁信息管理和运输优化方面具有重要应用价值，有利于提升城市的交运能力和智慧水平。

在地铁客流大数据分析与处理过程中，如何通过站点类型识别引导优化城市空间结构布局、有效挖掘客流时空分布和乘客出行规律、准确进行客流量状态监测等问题亟待解决。本项目面向地铁客流大数据开展了统计机器学习的一些关键技术研究，主要包括：考虑地铁复杂网络结构和乘客出行的规律多样性以及时空分布异质性，通过非监督学习挖掘客流数据潜在时空模式；构建针对高维稀疏矩阵的聚类算法，解决了聚类稳定性问题，从而提升了聚类结果的有效性及可解释性；在此基础上，对带有时间刻度和位置信息的非结构化数据进行分析和建模，通过构建站点客流模型来甄别客流影响因素，并提出了面向短时客流量预测的预测方法框架。在项目执行过程中，以第一/通讯作者发表SCI论文6篇，分别收录在《Physica A: Statistical Mechanics and its Applications》、《Transportation》、《IEEE Intelligent Transportation Systems Magazine》等领域权威期刊上；多次受邀参加国内外学术会议并作报告。