Nonlinear Traffic Prediction as a Matrix Completion Problem with Ensemble Learning

Li, Wenqing; Yang, Chuhan; Jabari, Saif Eddin · 2020 · OpenAlex-citations

DOI: 10.1287/trsc.2021.1086

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of short-term, high-resolution traffic prediction for signalized traffic operations management. Traditional forecasting methods typically rely on aggregated data with intervals of at least five minutes, which loses critical non-stationary features and reduces accuracy for adaptive control systems that operate on a second-by-second basis. The authors propose a novel framework that models this prediction task as a matrix completion problem, leveraging the sparsity and large volume of high-resolution sensor data. The primary motivation is to enable real-time, accurate predictions of sensor states (vehicle presence) to support coordinated adaptive traffic controllers, which require precise short-term forecasts of vehicle arrivals. The methodology formulates the prediction of future detector states as completing a low-rank matrix constructed from historical inputs and outputs. To handle the nonlinear relationships inherent in traffic data, the authors employ kernel regression, specifically using a radial basis function with periodic patterns (RBFP) to capture temporal correlations and signal cycle periodicities. This transforms the problem into a kernelized matrix completion task. To solve this efficiently, the authors develop a block-coordinate descent algorithm, which they prove converges in sub-linear time to a block coordinate-wise optimizer. Furthermore, they introduce an ensemble learning approach (adaptive boosting) that aggregates predictions from models trained on data from past days. This ensemble strategy aims to reduce training error to arbitrary thresholds while capturing recurring periodic trends, offering interpretability often lacking in deep learning models. The study evaluates the proposed method through theoretical analysis and empirical testing using both simulated data and a real-world high-resolution dataset from Abu Dhabi, UAE. The experimental design compares the proposed matrix completion approach against state-of-the-art algorithms, including parametric time series models (ARIMA, VARIMA) and non-parametric methods like support vector regression and various neural network architectures. The results demonstrate that the proposed method outperforms these existing techniques in prediction accuracy. The ensemble learning component successfully reduces training errors and effectively exploits periodic patterns in the data. The significance of this work lies in its ability to handle the computational challenges of high-resolution traffic data while maintaining high predictive accuracy. By framing prediction as a matrix completion problem, the approach provides a parsimonious solution that adapts to streaming data without the bias introduced by data aggregation. The sub-linear convergence of the algorithm ensures computational feasibility for real-time applications, such as adaptive signal control. Additionally, the ensemble method offers a transparent way to leverage historical periodicity, addressing the need for reliable, fine-grained predictions in complex urban traffic networks. This contributes a robust, data-driven tool for improving traffic efficiency and safety through advanced adaptive control systems.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• telematics crash prediction