zorrito 0.2.0

A lean reimplementation of Zorro: valid, sparse, stable explanations for PyTorch Geometric GNNs, with graph-level and regression support.

zorrito

A lean, modernized reimplementation of Zorro — valid, sparse, stable explanations for PyTorch Geometric GNNs.

zorrito keeps the greedy core of the paper extends the following

• Modern Python and Pytorch Geometric support
• Graph-level explanations, not just node-level
• Regression support via a tolerance band, with optional one-sided (direction-aware) match functions
• Configurable noise — whole-row sampling and dataset-wide pools, so perturbed inputs stay on the data manifold (matters a lot for categorical / one-hot features like atom types)

---

Install

pip install zorrito

zorrito depends on torch>=2.0 and torch-geometric>=2.4. Install those according to the official PyG instructions for your CUDA/PyTorch combination first.

For development:

git clone https://github.com/your-user/zorrito.git
cd zorrito
uv venv --python 3.11
uv pip install -e ".[dev]"
pytest tests/

See DEVELOP.md for the release / publishing checklist.

Quick start

Node classification (Cora-style)

import torch
from zorrito import Zorro

# `model` is any nn.Module that implements forward(x, edge_index)
explainer = Zorro(
    model=model,
    task="node",
    samples=100,
    top_k=10,
    seed=42,
)
explanations = explainer.explain(
    x=data.x,
    edge_index=data.edge_index,
    node_idx=10,
    fidelity_threshold=0.85,
)
expl = explanations[0]
print(expl.fidelity)                       # final RDT-Fidelity reached
print(expl.selected_node_indices())        # indices into the subgraph
print(expl.selected_feature_indices())     # indices into feature columns
print(expl.subgraph_nodes)                 # mapping back to the original graph

Graph classification (MUTAG-style)

from zorrito import Zorro

# `model` is any nn.Module that implements forward(x, edge_index, batch)
explainer = Zorro(
    model=model,
    task="graph",
    noise_mode="row",                # recommended for categorical features
    noise_pool=dataset_atom_pool,    # dataset-wide pool of valid atoms
    samples=100,
    seed=42,
)
explanations = explainer.explain(
    x=graph.x,
    edge_index=graph.edge_index,
    fidelity_threshold=0.95,
)

Regression with a one-sided direction

explainer = Zorro(
    model=model,
    task="graph",
    objective="regression",
    tolerance=0.4,                   # ε for the tolerance band
    direction="down",                # which atoms prevent the prediction
                                     # from FALLING below ref - ε
    noise_mode="row",
    noise_pool=dataset_atom_pool,
    seed=42,
)
explanations = explainer.explain(
    x=molecule.x,
    edge_index=molecule.edge_index,
    fidelity_threshold=0.95,
)

Key concepts

RDT-Fidelity. For a candidate explanation $\mathcal{S} = (V_s, F_s)$:

$$\mathcal{F}(\mathcal{S}) = \mathbb{E}{Z\sim\mathcal{N}}\big[,\mathrm{match}(\Phi(Y\mathcal{S}), \Phi(X)),\big]$$

where $Y_\mathcal{S}$ is the original input with the unselected entries replaced by random draws $Z$ from a noise distribution. zorrito estimates this with samples Monte-Carlo trials.

Match functions (the per-trial yes/no test):

objective	direction	match condition
classification	n/a	argmax(Φ(Y_S)) == argmax(Φ(X))
regression	both	`
regression	up	Φ(Y_S) ≤ Φ(X) + tolerance
regression	down	Φ(Y_S) ≥ Φ(X) − tolerance

For regression, direction="up" selects atoms that prevent the prediction from rising, and direction="down" selects atoms that prevent it from falling. This is often more informative than the symmetric tolerance band when the two sides have different chemistry.

Greedy algorithm. Starting from the empty selection, at each iteration zorrito evaluates the top-K candidate nodes and top-K candidate features (the "k" in top_k), adds the single element with the largest fidelity gain, and stops when fidelity exceeds fidelity_threshold. Then optionally re-runs to enumerate disjoint alternative explanations (max_explanations).

Noise distribution. The original paper draws each noise cell independently from its column's empirical distribution. zorrito keeps that as the default for node tasks (continuous features), and switches to whole-row sampling by default for graph tasks (categorical / one-hot features stay valid). The pool the rows are drawn from can be set independently via noise_pool.

Configuration cheat sheet

Zorro(
    model,
    task               = "node" | "graph",
    objective          = "classification" | "regression",
    select             = "both" | "nodes_only" | "features_only",
    direction          = "both" | "up" | "down",     # regression only
    noise_pool         = None | torch.Tensor,        # default: x from explain()
    noise_mode         = "column" | "row",           # default: column (node), row (graph)
    device             = "cpu" | "cuda",
    samples            = 100,
    top_k              = 10,
    tolerance          = 0.0,
    num_hops           = 2,                          # node-task subgraph radius
    seed               = None,
    log                = False,
)

explainer.explain(
    x,
    edge_index,
    node_idx           = None,                       # required for task="node"
    batch              = None,
    fidelity_threshold = 0.85,
    max_explanations   = 1,
)

The return is a list[Explanation], where each Explanation exposes node_mask, feature_mask, fidelity, trace, and (for node tasks) subgraph_nodes. The masks are boolean tensors; use selected_node_indices() / selected_feature_indices() to convert.

Notable departures from the original paper / reference implementation

• fidelity_threshold=0.85 is the value $\tau$ from the paper. The original code uses tau=0.15 (which is 1 − τ_paper)
• max_explanations replaces recursion_depth.
• Explanations are returned as a single Explanation dataclass with boolean masks, instead of nested numpy arrays.
• The package targets modern Python (>=3.10) and PyG (>=2.4).

Citing the algorithm

The algorithm is from the original Zorro paper:

@article{funke2021zorro,
  title   = {Zorro: Valid, Sparse, and Stable Explanations in Graph Neural Networks},
  author  = {Funke, Thorben and Khosla, Megha and Rathee, Mandeep and Anand, Avishek},
  journal = {arXiv preprint arXiv:2105.08621},
  year    = {2021}
}

License

MIT — see LICENSE.