A research program on structural lower bounds for graph-supervised learning, instantiated empirically on urban mobility, bike-share demand and materials informatics, and shipped as open, reproducible software and datasets.
We develop a universal spectral lower bound on the generalisation error of any graph-supervised learner, depending only on the graph Laplacian, the target signal and the learner's reachable feature subspace, and instantiate it across bike-share, mobility and materials data. Drafts mature privately and are opened as preprints and public repositories as they go out; alongside the theory we maintain open, research-grade software and datasets. The repositories below are the ones currently public.
| Repository | What it is |
|---|---|
| gbfs-toolkit | The research-grade Python library for GBFS bike-share feeds: ingestion into a stable, version-independent data model, the semantic A1-A7 static and D1-D3 dynamic quality audit, and descriptive spatial and temporal analytics. Installable from PyPI (pip install gbfs-toolkit), with versioned documentation and a Zenodo DOI. |
| gbfs-audit-catalogue | A certified, anomaly-flagged reference dataset for bike-sharing station feeds (1,509 GBFS systems across 48 countries; 46-column schema), built on the gbfs-toolkit audit pipeline, with a project page. Dataset v1.0 released with a Zenodo DOI. |
| sbf-lean | Machine-checked Lean 4 / Mathlib proof of the universal spectral lower bound (Theorem 1) and the deterministic ERM-oracle core of Theorem 2, with zero sorry. |
| od-lean | Machine-checked Lean 4 / Mathlib proof of the GBFS origin-destination identifiability bounds (the q⁻¹ Cramér-Rao limit, the bias-structure projection, and the collection-horizon law), with zero sorry and a SymPy cross-check. |
Formal verification. A distinctive thread of the program is that its core theorems are machine-checked: the Lean developments above build with zero
sorry, depending only on Lean's three standard axioms. A green CI run is the public certificate behind each claim.
Each public repository ships a pinned environment, an MIT LICENSE (data products inherit
upstream licences, see LICENSE-DATA in the catalogue) and a CITATION.cff; gbfs-toolkit
and gbfs-audit-catalogue releases are archived on Zenodo with a DOI. The software is
continuously tested across Python versions and operating systems, type-checked, and
published to PyPI through a test-gated release pipeline with build provenance attestation.
- Rohan Fossé · CESI École d'Ingénieurs, Montpellier · rohanfosse.fr · rfosse@cesi.fr · ORCID
- Gaël Pallares · CESI LINEACT (EA 7527) · ORCID
Affiliated with CESI LINEACT (EA 7527), Montpellier, France.
@misc{cyclingDataLab,
author = {Foss\'e, Rohan and Pallares, Ga\"el},
title = {{cycling-data-lab}: a research program on structural lower bounds
for graph-supervised learning},
year = {2026},
howpublished = {\url{https://github.com/cycling-data-lab}}
}One spectral lower bound, several empirical domains, with machine-checked proofs and open, reproducible artifacts.