In brief
This study asks whether a graph forecasting model can learn better node representations by training across several related datasets instead of learning every representation from scratch.
Contribution
A subgraph-based multi-dataset training procedure, followed by embedding transfer and fine-tuning, together with an explainability analysis of the representations.
Key finding
Transferred embeddings improved prediction accuracy on several traffic datasets while remaining competitive on the others.
Models and data
Distinctive research elements
Models and methods
- Multi-dataset graph training
- Transferable node embeddings
Data and evaluation
- Traffic forecasting graphs
- Generated subgraph dataset
Author-written overview
Research overview
Graph forecasting models learn an embedding for each node, but those representations are often trained independently for every dataset. This study examines whether related spatiotemporal graphs can instead contribute to a shared representation. It constructs a training collection from subgraphs of several traffic datasets, trains an adapted model across those graphs and then transfers and fine-tunes the resulting embeddings on target forecasting tasks. An explainability analysis is used to examine how multi-dataset training changes the learned node space. The transferred embeddings improve prediction accuracy on several validation datasets and remain competitive on the others, supporting multi-graph training as a practical route to stronger reusable representations.
Cite this work
BibTeX
@article{garciasiguenza2026multi,
title = {Multi-Dataset Training for Improved Accuracy in Spatio-Temporal Problems: An Explainable Analysis},
author = {Javier García-Sigüenza and Alberto Real-Fernández and Faraón Llorens-Largo and Rafael Molina-Carmona and Marc Semper},
journal = {Mathematics},
year = {2026},
volume = {14},
number = {5},
eid = {908},
doi = {10.3390/math14050908}
}Always check the publisher record for final volume, issue and page information before citing.