基于图卷积网络的交通预测方法研究
张奇
2021-05
Pages116
Subtype博士
Abstract

作为智能交通系统的重要组成部分,交通预测是众多交通应用的基础。性能良好的交通预测系统可以提供及时精确的交通信息,有助于缓解拥堵,提高出行效率,减少能耗和污染。由于其重要意义,交通预测受到了学术界和工业界的高度重视。


由于交通路网的不规则性,嵌入在交通路网上的交通数据本质上是一种图结构数据。同时,交通数据有很多特殊性质:平稳性的统计特性、复杂的局部结构、动态性和全局关联性。针对这些特性,本文设计了多种有效的图卷积方法来处理交通预测问题。本文的贡献包含以下几个方面:

针对交通数据非平稳性的特点,提出了一种基于核加权图卷积神经网络的交通预测方法。该方法的核心思想是将多个卷积核的加权求和结果作为实际参与运算的卷积核,并通过调整卷积核的加权系数来放松权值共享约束。具体地,核加权图卷积同时学习多个候选卷积核以及随位置变化的卷积核线性组合系数。由于在不同位置实际使用了不同的卷积核参数,核加权图卷积放松了继承于经典卷积神经网络的权值共享约束,从而更适合处理非平稳的交通数据。在数学上,本文证明了所提方法可以被视为局部连接网络的低秩近似。实验结果验证了所提方法的有效性。


针对交通数据具有复杂局部结构的特点,提出一种基于局部静态结构学习的交通预测方法。具体地,该方法通过学习一个局部静态图结构来掌握相邻节点之间的关系,从而更好的捕获交通数据中的空间信息。技术上,引入了图结构系数并将其作为可学习的参数直接参与卷积运算,从而有助于所提方法的卷积核聚焦于邻域内的重要节点。此外,通过将图结构系数与卷积核参数相结合,所提方法放松了权值共享的约束,能较好地处理非平稳的图结构数据。在多个数据集上充分的实验评估证明了所提方法的有效性。

针对交通数据动态性和全局关联性的特点,提出了一种基于结构学习卷积神经网络的交通预测方法。首先,提出了结构学习卷积这一泛化性较强的图卷积框架,多种现有的图卷积方法可视为其特例。沿着这一技术路线,本文构建了两个结构学习模块以分别捕获全局和局部结构。每个模块分别包含一个静态结构学习项(用于学习所有样本的共享结构),和一个动态结构学习项(用于学习每个样本的独特结构)。模块中的每一项均可视为结构学习卷积的一种特殊实例化。此外,所提方法中加入了伪三维卷积模块来捕获交通数据中的时间依赖关系。该方法在六个交通数据集上进行评估,大量的对比实验表明所提出的方法优于当前主流方法。

Other Abstract


As an indispensable part of the Intelligent Traffic System (ITS), traffic forecasting is the foundation of many traffic applications. A traffic forecasting system with good performance can provide timely and accurate forward-looking traffic information, which is helpful to ease congestion, improve travel efficiency as well as reduce energy consumption and pollution. Due to its usage, traffic forecasting has received great attention both in the academic and industrial communities.

Because of the complicated topological structure of traffic networks, the traffic data embedded in traffic networks is inherently a kind of graph-structured data.Meanwhile, traffic data has many special characteristics. It has non-stationary statistical property and complex local structure. In addition, the relationships between nodes change dynamically, and are related to each other globally. Considering these characteristics, this dissertation designs several effective graph convolution methods to deal with traffic forecasting problems. The contributions of this dissertation include the following aspects:


To consider the non-stationary characteristics of traffic data, a traffic forecasting method based on Kernel-Weighted Graph Convolutional Neural network (KWGCN) is proposed. The core idea behind this method is to take the weighted sum result of  multiple candidate  convolution kernels as the actually convolutional kernel that participates in the convolution operation, then relax the weight sharing constraint by  adjusting the  linear combination coefficients.Specifically, KWGCN learns simultaneously  multiple candidate convolution kernels and their linear combination coefficients that vary with locations.KWGCN actually uses different convolution kernel parameters at different locations,thus it relaxes the weight sharing constraint inherited from the classical convolutional neural network and it is more  suitable to process non-stationary traffic data. Mathematically, we prove that the proposed method can be regarded as a low-rank approximation of locally-connected networks.In addition,  several traffic data sets are constructed in this dissertation to verify the validity of the proposed model. Experimental results on both public datasets and private datasets demonstrate the effectiveness of the proposed method.


To consider the complex local structure of traffic data, a traffic forecasting method based on local static structure learning is proposed. Specifically, by learning a local static graph structure to measure the correlation degrees of adjacent nodes, the proposed method can better capture the spatial information in traffic data. Technically, graph structure parameters are introduced as learnable parameters, which are directly involved in convolution operation and  help the filters of the proposed method to focus on the important nodes in the neighborhood. Meanwhile, by learning the graph structure parameters and kernel weights, our method  relaxes the restriction of weight sharing and  handles the graph-structured data of non-stationarity efficiently. Comprehensive experimental evaluations on several datasets verify the effectiveness of the proposed method.


To consider the  dynamic and global correlation of traffic data, a traffic forecasting method based on structural learning convolutional neural network is proposed. Firstly, we proposed Structure Learning Convolution (SLC). As a  general framework, some existing graph convolution methods  can be regarded as a special case of SLC. Then, along this technical line, two SLC modules are proposed to capture the global and local structures, respectively. Each module contains a static structure learning term (to learn the shared structure of all samples) and a dynamic structure learning item (to learn the unique structure of each sample), respectively. Each term in the module can be regarded as a special instantiation of SLC. Additionally,  Pseudo three-dimensional convolution networks are combined with the proposed method to capture the temporal dependencies in traffic data. Extensively comparative experiments on six real-world datasets demonstrate the proposed approach significantly outperforms the state-of-the-art ones.

Keyword交通预测 图卷积网络 时空数据挖掘 深度学习
Language中文
Sub direction classification机器学习
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/45053
Collection多模态人工智能系统全国重点实验室_先进时空数据分析与学习
Recommended Citation
GB/T 7714
张奇. 基于图卷积网络的交通预测方法研究[D]. Institute of Automation, Chinese Academy of Sciences. Institute of Automation, Chinese Academy of Sciences,2021.
Files in This Item:
File Name/Size DocType Version Access License
Thesis_4_3.pdf(12681KB)学位论文 开放获取CC BY-NC-SA
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[张奇]'s Articles
Baidu academic
Similar articles in Baidu academic
[张奇]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[张奇]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.