Yu, Taeyoung (The University of Queensland), Kim, Jiwon (The University of Queensland), Lone, Soban Nasir (Technical University of Munich), Abouelela, Mohamed (Technical University of Munich), Antoniou, Constantinos (Technical University of Munich)

Scalable Traffic Prediction Via Data Reduction with Coreset Selection

Scheduled for presentation during the Regular Session "S30b-Intelligent Modeling and Prediction of Traffic Dynamics" (TH-EA-T30), Thursday, November 20, 2025, 14:50−14:50, Gold Coast

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 18, 2025

Abstract

Deep learning models for large-scale traffic forecasting require significant computation for training massive datasets. Coreset selection, which selects small representative data subsets, can reduce computational burden while maintaining prediction accuracy. We comprehensively evaluated eight model-free coreset selection methods for short-term traffic flow prediction, using a Spatio-Temporal Graph Convolutional Network (STGCN) model on the PEMS04 and PEMS07 datasets. Our experiments show that predictive performance comparable to the full dataset model can be maintained with only 30-50% of the data, without a significant performance drop. Furthermore, using 50-90% of the data sometimes leads to slightly better performance than training on the entire dataset. The clustering-based approach achieved the highest overall accuracy, suggesting that clustering can effectively select representative samples while filtering out noisy samples. Our analysis indicates that feature space coverage is connected to coreset selection performance and that temporal diversity can be utilised to enhance this coverage. Coresets with strong temporal bias performed poorly, failing to achieve broad feature-space coverage.