Reproduce “ETA Prediction with Graph Neural Networks in Google Maps” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Austin Derrow-Pinion, Jennifer She, David Wong, Oliver Lange, Todd Hester, Luis Perez, Marc Nunkesser, Seongjae Lee and 9 more
Google Maps has integrated a Graph Neural Network (GNN) model for estimated time of arrival (ETA) prediction, leveraging road network topology to anticipate future traffic conditions more accurately. This system reduces "negative ETA outcomes" by over 40% in certain cities compared to prior methods.
Key insights
- Graph Neural Networks (GNNs) effectively model complex spatiotemporal interactions within road networks, outperforming traditional and simpler deep learning baselines for ETA prediction.
- Training stability is paramount for production-grade GNNs; techniques like MetaGradients and Exponential Moving Average (EMA) of parameters are critical for reducing variance and ensuring robustness.
- Learnable segment and supersegment embeddings, coupled with a combination of supersegment, segment, and cumulative segment auxiliary losses, significantly enhance model performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.