Federated_Graph_Learning

A repository for synthetic subgraph-detection benchmarking and PNA baselines on directed multigraphs.

Synthetic Graph Generation

This repository includes a synthetic subgraph-detection dataset used for benchmarking graph models for the pattern detection task. The graphs and labels are generated following the pseudocode/configurations described in Provably Powerful Graph Neural Networks for Directed Multigraphs (Egressy et al.).

What’s Included

• Three splits: train, val, test

• Saved as PyTorch tensors under ./data/:

  • train.pt, val.pt, test.pt objects with node-level labels
  • y_sums.csv — per-split counts of positive labels per sub-task

• Per-split label percentages and mean across splits are stored under ./results/metrics/, useful to sanity-check against the paper’s reported marginals

Label Tasks

Each node is labeled for the presence of the following patterns (11 sub-tasks):

• deg_in > 3
• deg_out > 3
• fan_in > 3
• fan_out > 3
• cycle2
• cycle3
• cycle4
• cycle5
• cycle6
• scatter_gather
• biclique

Reproducibility

Graph instances are reproducible. A single BASE_SEED deterministically derives distinct seeds for each split (train/val/test), ensuring:

• different graphs within a run for the splits,
• identical graphs across runs with the same BASE_SEED.

Default Generation Settings

The default config (see the generator script scripts/generate_synthetic.py) follows the paper’s setup:

• Nodes n = 8192
• Average degree d = 6
• Radius parameter r = 11.1
• Directed multigraphs (for directed cycles)
• Generator: “chordal” / random-circulant-like
• One connected component per split (prevents data leakage)

How to Generate

From the repo root:

# Generate graphs and labels
python3 -m scripts.generate_synthetic

After step (1), you’ll find train.pt, val.pt, test.pt, and y_sums.csv under ./data/. The label_percentages.csv will be saved under ./results/metrics/.

---

Principal Neighborhood Aggregation (PNA)

PNA model is implemented by following the model architecture described in Principal Neighbourhood Aggregation for Graph Nets (Corso et al.).

Train Baseline PNA

From the repo root:

# Train and test PNA model on the generated graph data
python3 -m scripts.train_pna_baseline

Train PNA with Reverse Message Passing

Training the PNA model with Reverse Message Passing, which uses a Heterogeneous Graph and Ego IDs to detect the fraud patterns listed under Label Tasks.

From the repo root:

# Train and test enhanced PNA model on the generated graph data
python3 -m scripts.train_pna_reverse_mp_with_ego

Hyperparameter Tuning

From the repo root:

python3 -m scripts.pna_hyperparameter_tuning