Tensor decomposition for spatial-temporal traffic flow prediction with sparse data

Urban transport traffic surveillance is of great importance for public traffic control and personal travel path planning. Effective and efficient traffic flow prediction is helpful to optimize these real applications. The main challenge of traffic flow prediction is the data sparsity problem, meanin...

Full description

Saved in:
Bibliographic Details
Main Authors: Yang, Funing, Liu, Guoliang, Huang, Liping, Chin, Cheng Siong
Other Authors: School of Electrical and Electronic Engineering
Format: Article
Language:English
Published: 2021
Subjects:
Online Access:https://hdl.handle.net/10356/145688
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Nanyang Technological University
Language: English
Description
Summary:Urban transport traffic surveillance is of great importance for public traffic control and personal travel path planning. Effective and efficient traffic flow prediction is helpful to optimize these real applications. The main challenge of traffic flow prediction is the data sparsity problem, meaning that traffic flow on some roads or of certain periods cannot be monitored. This paper presents a transport traffic prediction method that leverages the spatial and temporal correlation of transportation traffic to tackle this problem. We first propose to model the traffic flow using a fourth-order tensor, which incorporates the location, the time of day, the day of the week, and the week of the month. Based on the constructed traffic flow tensor, we either propose a model to estimate the correlation in each dimension of the tensor. Furthermore, we utilize the gradient descent strategy to design a traffic flow prediction algorithm that is capable of tackling the data sparsity problem from the spatial and temporal perspectives of the traffic pattern. To validate the proposed traffic prediction method, case studies using real-work datasets are constructed, and the results demonstrate that the prediction accuracy of our proposed method outperforms the baselines. The accuracy decreases the least with the percentage of missing data increasing, including the situation of data being missing on neighboring roads in one or continuous multi-days. This certifies that the proposed prediction method can be utilized for sparse data-based transportation traffic surveillance.