From Stops to Graphs
Hands-On Spatio-Temporal Learning for Public Transportation

01 · Overview

Urban public transport systems produce massive spatiotemporal data streams combining stops, routes, vehicles, and passenger flows. The SUNT dataset captures one year of operations in Salvador, Brazil, at high temporal resolution, enabling realistic modeling of operational dynamics.

This hands-on tutorial teaches how to learn from spatiotemporal public transportation data using deep neural networks, foundation models, and graph neural networks. Using our dataset, participants will build a preprocessing pipeline, train, and compare temporal models such as LSTM, GRU, TCN, and Transformers, and model spatial structure with GCN, GraphSAGE, and GAT. We then couple space and time with spatio-temporal architectures and evaluate both regression and classification tasks. All steps are reproducible, and attendees leave with notebooks, code, and clear practices for validation and error analysis.

02 · Learning objectives

By the end of this tutorial, participants will be able to:

• Understand the structure and main components of the SUNT dataset.
• Build stop-level and line-level graphs from spatiotemporal transit data.
• Implement baseline GNN models for prediction tasks on public transport networks.
• Compare GNNs with standard temporal models in this domain.
• Identify open challenges and research opportunities in GNNs for mobility.

03 · Schedule

The tutorial is organized as a 4-hour, hands-on session that progressively moves from understanding the SUNT dataset to advanced GNN models and open research directions.

04 · Prerequisites

We assume that participants are comfortable with:

• Python programming (basic scripting, functions, and notebooks).
• Fundamentals of machine learning and deep learning.
• Basic graph terminology (nodes, edges, adjacency, neighborhoods).

Helpful, but not strictly required:

• Previous experience with PyTorch or similar deep learning frameworks.
• Some familiarity with time-series or spatiotemporal data.

05 · Software setup

To follow the hands-on part, we recommend that participants bring a laptop with the following tools installed:

• Python 3.10+ (Anaconda/Miniconda is recommended).
• Jupyter Notebook or JupyterLab.
• PyTorch and a GNN library (e.g., PyTorch Geometric or DGL).

We will provide a ready-to-use environment file and setup instructions in the repository:

environment.yml or requirements.txt will be available at
TBD

For participants without a local setup, we will also share links to Google Colab notebooks that can be run directly in the browser.

06 · Materials

The following materials will be made available before and during the tutorial:

• Slides PDF Tutorial slides summarizing the main concepts and case studies.
• Notebooks Colab Step-by-step notebooks for data loading, graph construction, and GNN training.
• Code Reproducible PyTorch code for baseline models and experiments.
• Dataset Link to the official SUNT dataset and documentation.

Dataset paper:
SUNT: A Landmark Spatiotemporal Dataset for Public Transportation

07 · Instructors

08 · Contact

If you have any questions about the tutorial, prerequisites, or materials, please feel free to contact the organizers:

• Email: ricardoar [at] ufba [dot] br
• GitHub: https://labia.ufba.br

We look forward to seeing you at WCCI / IJCNN 2026 and exploring GNNs for public transportation modeling together.