Enhanced Application of Principal Component Analysis in Machine Learning for Imputation of Missing Traffic Data

Yoon-Young Choi,Seungmo Kang,Dong-Kyu Kim,Heeseung Shon,Young-Ji Byon

doi:10.3390/app9102149

Abstract

Missing value imputation approaches have been widely used to support and maintain the quality of traffic data. Although the spatiotemporal dependency-based approaches can improve the imputation performance for large and continuous missing patterns, additionally considering traffic states can lead to more reliable results. In order to improve the imputation performances further, a section-based approach is also needed. This study proposes a novel approach for identifying traffic-states of different spots of road sections that comprise, namely, a section-based traffic state (SBTS), and determining their spatiotemporal dependencies customized for each SBTS, for missing value imputations. A principal component analysis (PCA) was employed, and angles obtained from the first principal component were used to identify the SBTSs. The pre-processing was combined with a support vector machine for developing the imputation model. It was found that the segmentation of the SBTS using the angles and considering the spatiotemporal dependency for each state by the proposed approach outperformed other existing models.

Highlights

A traffic state of congestion generally arises at the sites that have traffic volume exceeding the associated road capacity
The objective of this study is to propose a new approach for the imputation of missing data by identifying section-based traffic state (SBTS) of a target location, and determining tempo-spatial dependencies customized for each SBTS, with data at different time periods from upstream/downstream traffic detectors in the vicinity of the target
This study proposes a novel approach of imputation of missing traffic data by identifying SBTS

Summary

Introduction

A traffic state of congestion generally arises at the sites that have traffic volume exceeding the associated road capacity. In order to address this problem, departments of transportation (DOTs) and other authorities spend significant portions of their budgets on intelligent transportation system (ITS) applications to monitor traffic flows and manage congestion-related issues. Data collected from stationary detectors including loop detectors are most widely used for monitoring the traffic conditions. The ITS applications ideally need complete and continuous streams of traffic data in order to function properly, in reality, significant portions of the collected data from the loop detectors are often missing, causing flaws potentially resulting in under or overshooting errors with existing prediction models for ITS applications [1,2,3,4]. Qu et al [5] report that roughly 10% of daily traffic volume data is missing in Beijing, China, mainly due to malfunctions of detectors.

Objectives

Methods

Findings

Conclusion