FR-MTL: Traffic Flow Prediction Using Fused Ridge Denoising and Multi-Task Learning

• Published in: Journal of advanced computational intelligence and intelligent informatics 2020-12, Vol.24 (7), p.829-836
• Main Authors: Yang, Di, Qiu, Ningjia, Wang, Peng, Yang, Huamin
• Format: Article
• Language: English
• Subjects: Coefficient of variation; Correlation analysis; Intelligent transportation systems; Learning; Missing data; Noise reduction; Noise sensitivity; Prediction models; Random noise; Sparsity; Traffic flow; Traffic information; Traffic models
• ISSN: 1343-0130; 1883-8014
• DOI: 10.20965/jaciii.2020.p0829

Traffic flow prediction is one of the fundamental components in Intelligent Transportation Systems (ITS). Many traffic flow prediction models have been developed, but with limitation of noise sensitivity, which will result in poor generalization. Fused Lasso, also known as total variation denoising, penalizes L 1 -norm on the model coefficients and pairwise differences between neighboring coefficients, has been widely used to analyze highly correlated features with a natural order, as is the case with traffic flow. It denoises data by encouraging both sparsity of coefficients and their differences, and estimates the coefficients of highly correlated variables to be equal to each other. However, for traffic data, the same coefficients will lead to overexpression of features, and losing the trend of time series of traffic flow. In this work, we propose a Fused Ridge multi-task learning (FR-MTL) model for multi-road traffic flow prediction. It introduces Fused Ridge for traffic data denoising, imposes penalty on L 2 -norm of the coefficients and their differences. The penalty of L 2 -norm proportionally shrinks coefficients, and generates smooth coefficient vectors with non-sparsity. It has both capability of trend preservation and denoising. In addition, we jointly consider multi-task learning (MTL) for training shared spatiotemporal information among traffic roads. The experiments on real traffic data show the advantages of the proposed model over other four regularized baseline models, and on traffic data with Gaussian noise and missing data, the FR-MTL model demonstrates potential and promising capability with satisfying accuracy and effectiveness.