TopoGeoML

Topology-aware machine learning tools for persistent-homology features, differentiable topology losses, simplicial and Hodge operators, signal analysis, and reproducible evaluation.

TopoGeoML is a Python scientific-software library for cases where connected components, loops, persistence, simplicial structure, or topology through training are part of the signal you want to measure or constrain. It provides a scikit-learn feature pipeline, PyTorch topology losses, simplicial and Hodge primitives, topology features for signals, embedding diagnostics, and a reproducible research harness.

The repository also contains a preregistered graph-classification investigation. That study is evidence about where one family of topological methods did and did not help. It is not the definition of the library.

Where TopoGeoML fits

TopoGeoML is most useful when topology is a plausible source of signal or a desired structural constraint and you want a small, inspectable pipeline around it. It is not intended to replace broad computational-topology libraries, general graph-learning frameworks, or specialized high-throughput GPU persistence implementations.

What the graph study established

The graph study asks whether Hodge-Laplacian propagation itself provides a unique classification advantage. Across the tested matched-capacity configurations, no unique L_0 Hodge advantage is supported. In H008c on NCI1, a normalized-adjacency arm using the same external identity-skip architecture as Hodge reaches 0.629 versus 0.609 for Hodge. H010 then finds a significant adjacency advantage on MUTAG, no significant operator difference on PROTEINS, and the same favorable adjacency direction on NCI1.

H008c also shows that the tested external-residual adjacency formulation recovers performance after normalization with an internal self contribution does not. That causal statement is intentionally scoped. gin-normalised places a trainable self term inside the affine/nonlinear update, whereas gin-residual adds an identity skip outside the activation. The result identifies the successful tested self-path formulation; it does not prove that residual connections alone are the sole mechanism in arbitrary architectures.

There is a narrow positive result. On NCI1, hodge-mp-residual outperforms the matched-capacity MLP baseline by a median 8.6 percentage points within the H003 comparison family (p_BH = 4.83e-3). Later operator ablations show that the improvement is not unique to the Hodge Laplacian. The former investigation-wide 59-comparison and 76-entry sensitivity analysis remains withdrawn as a current claim because it included invalidated H009 comparisons and has not been regenerated from the validated comparison set.

The sheaf question is now resolved by a corrective replication. The historical H009 run was invalidated by implementation audit (its arm did not guarantee the stated cellular-sheaf Laplacian and missed the capacity tolerance) and remains provenance only. The repaired, invariant-tested sheaf-residual 2.0.0 operator was rerun under the preregistered H009-R protocol: the learned scalar sheaf construction beats the matched MLP (p_BH = 4.73e-3), shows no significant difference from the fixed Hodge operator (p_BH = 0.428, not an equivalence claim), and sits below the matched normalized-adjacency arm without crossing H41's strict 0.01 falsification threshold (p_BH = 0.0342, inconclusive). No learned-sheaf advantage over the fixed operator was detected in this regime.

The higher-order question also remains open. H011's NCI1 L_1 arm does not significantly outperform MLP and is tested on a dataset where 96% of graphs have no triangles. The triangle-rich COLLAB follow-up H011b has only a one-seed directional smoke result so far.

This matters for users. TopoGeoML does not present L_0 Hodge propagation as a universal GNN upgrade. The negative, inconclusive, and invalidated results remove unjustified use cases while leaving the reusable topology features, losses, diagnostics, signal tools, and higher-order research primitives intact.

See STATUS.md, docs/STATISTICAL_SUMMARY.md, LEADERBOARD.md, and docs/CLAIMS_TO_EVIDENCE.md for the current evidence record. The older docs/RESEARCH_REPORT.md is explicitly preserved as the Version 0.0.2 historical snapshot through H008c.

Status

TopoGeoML 0.0.7 is beta scientific software. Public APIs can change before 1.0.

Required CI enforces 100% line and 100% branch coverage on the importable topogeoml package under full dependencies. Mypy strict mode and ruff are enforced in CI. The benchmarks/ tree is separate research infrastructure and is not included in the package-coverage claim. The exact verified test and coverage snapshot is recorded in docs/CLAIMS_TO_EVIDENCE.md.

The project reports negative, null, exploratory, invalidated, and positive findings separately. A non-significant comparison is not treated as proof of equivalence. An equivalence claim would require a preregistered equivalence margin and an appropriate equivalence procedure.

Installation

Core topology features and data structures:

pip install topogeoml

Differentiable Vietoris-Rips persistence and Hodge layers:

pip install "topogeoml[torch]"

Differentiable cubical persistence and CubicalTopologyLoss require both PyTorch and GUDHI:

pip install "topogeoml[torch,tda]"

For development from source:

git clone https://github.com/smaniches/TopoGeoML.git
cd TopoGeoML
pip install -e ".[dev]"
pytest

Quick start

Persistent-homology features in scikit-learn

import numpy as np

from topogeoml import TopologyFeaturePipeline

rng = np.random.default_rng(42)
theta = np.linspace(0, 2 * np.pi, 50, endpoint=False)
t = np.linspace(-1, 1, 50)

circle = np.stack([np.cos(theta), np.sin(theta)], axis=1)
circle += 0.05 * rng.standard_normal((50, 2))

line = np.stack([t, np.zeros(50)], axis=1)
line += 0.05 * rng.standard_normal((50, 2))

pipe = TopologyFeaturePipeline(max_homology_dim=1, resolution=10)
features = pipe.fit_transform([circle, line])

print(features.shape)  # (2, 200)

TopologyFeaturePipeline is a standard scikit-learn transformer, so it can be placed inside a Pipeline and fit within each cross-validation fold.

Cubical topology loss for segmentation

import torch

from topogeoml.nn.cubical_diff_ph import CubicalTopologyLoss

# Example structural prior: one foreground connected component.
topo_loss = CubicalTopologyLoss(target_betti={0: 1}, invert=True)

pred = torch.rand(
    4, 1, 64, 64,
    dtype=torch.float64,
    requires_grad=True,
)

loss = topo_loss(pred)
loss.backward()

Use this as an auxiliary structural loss when the desired topology is known. The implementation is tested as a differentiable primitive; the repository does not claim that adding it improves every segmentation task.

Hodge message passing

import networkx as nx
import torch

from topogeoml import graph_to_clique_complex
from topogeoml.nn.hodge import build_hodge_layer_from_complex

complex_ = graph_to_clique_complex(nx.complete_graph(5), max_dim=2)
layer = build_hodge_layer_from_complex(
    complex_,
    k=0,
    in_features=16,
    out_features=8,
)

x = torch.randn(complex_.n_simplices(0), 16)
out = layer(x)
print(out.shape)  # torch.Size([5, 8])

At k=0, the normalized Hodge operator is the normalized graph Laplacian. On positive-degree vertices it has the familiar form I - D^{-1/2} A D^{-1/2}; isolated vertices follow the implementation's zero degree-support convention and keep zero normalized rows. It is not the standard Kipf-Welling GCN propagation operator. HodgeMessagePassing is a fixed-complex research building block, not a claim of state-of-the-art graph classification.

Implemented library surface

Torch-dependent neural symbols are imported from their subpackages so the core package remains importable without PyTorch.

Research harness

The benchmarks/, notebooks/, and hypothesis files belong to the source repository and support the empirical record. They are not the same thing as the installed topogeoml wheel.

A short Hodge smoke run from a source checkout is:

pip install -e ".[all]"
python -m benchmarks.hodge \
  --datasets mutag \
  --seeds 0 1 2 \
  --n-epochs 5

This exercises the real benchmark path but is not a reproduction of the full seeded claims. Exact reproduction commands for the preregistered studies are in REPRODUCING.md.

Evidence discipline

The repository separates software correctness from empirical claims.

• Package correctness is gated by tests, strict type checking, linting, dependency audit, and 100% line and branch coverage under the full-dependency package gate.
• Empirical claims point to seeded result artifacts and reproduction commands.
• Pairwise graph experiments use paired Wilcoxon tests and Benjamini-Hochberg correction within their declared families.
• The former investigation-wide 59/76 sensitivity analysis remains withdrawn; H009-R completes the validated comparison set, but no regenerated table is claimed yet.
• Non-significance is reported as no detected difference at the tested power, not as statistical equivalence.
• Smoke runs and exploratory diagnostics are not promoted to confirmatory claims.

What remains unproven

The public contract is intentionally narrower than the code surface.

• No claim says topology improves every ML task.
• No claim says Hodge propagation is generally better than a well-tuned GNN.
• The historical H009 result is not valid evidence about a cellular-sheaf Laplacian; the corrected H009-R replication is the valid record, and it shows no learned-sheaf advantage over the fixed Hodge operator at the tested configuration.
• No powered end-to-end result currently shows CubicalTopologyLoss improving medical-image segmentation.
• The H011b COLLAB L_1 experiment has a one-seed directional smoke result, but the full statistical run remains incomplete.
• The topology-divergence callback result is exploratory because it is floor-limited and lacks a non-overfitting negative control.
• Large-scale GPU persistence, variable-topology batched Hodge networks, and a broad computational-topology algorithm catalog are outside the current package scope.

See LIMITATIONS.md for the detailed engineering and scientific limits.

Citation

@software{maniches_topogeoml_2026,
  author  = {Maniches, Santiago},
  title   = {TopoGeoML: A Preregistered Investigation into Topology-Aware Graph Classification},
  year    = {2026},
  version = {0.0.7},
  doi     = {10.5281/zenodo.20365816},
  url     = {https://doi.org/10.5281/zenodo.20365816}
}

Machine-readable citation metadata is in CITATION.cff. DOI: 10.5281/zenodo.20365816.

License

MIT. See LICENSE.

Santiago Maniches (ORCID: 0009-0005-6480-1987).