kvirtue/fwa-gnn-rgcn

FWA GNN — Relational Graph Convolutional Network for Federal Fraud, Waste & Abuse Detection

A heterogeneous R-GCN trained on a synthetic federal grant/loan entity graph to classify contractor and applicant accounts as fraudulent or legitimate. This model is the Phase 2 artifact of the Federal FWA Research Project.

Model Description

Architecture: RGCN_FWA — a 2-layer Heterogeneous R-GCN built with PyTorch Geometric HeteroConv + RGCNConv.

The model operates on a heterogeneous knowledge graph with 7 node types and 15 edge types (including reverse edges). Classification is performed on Account nodes only.

Hyperparameter	Value
Hidden dim	128
Num layers	2
Num bases (RGCNConv)	30
Dropout	0.3
Optimizer	Adam, lr=1e-3 → 5e-4 → 2.5e-4 (step decay)
Loss	BCEWithLogitsLoss (pos_weight tuned for class imbalance)
Epochs trained	132

Graph Schema

Node Type	Count	Features
Account	17,135	19 CIO-derived features (see feature_stats.json)
Individual	21,448	3 dense features
Application	32,044	Learnable embedding
BankAccount	14,725	Learnable embedding
Address	28,479	Learnable embedding
Agency	12	Learnable embedding
PaymentEvent	55,345	Learnable embedding

Account node features (19-dim): sharing_peer_count, shared_address_count, app_count_6mo, inc_age_days, shared_address_edges, related_controlled_count, max_owner_biz_count, max_payment_count, max_total_amount, payment_hhi, one-hot business type (5), one-hot status (4).

Fraud Typologies Targeted

ID	Typology	Description
FT-SHELL	Shell company network	Shared agents/principals, reused bank accounts, overlapping ownership
FT-BIDRIG	Bid rigging / collusion	Coordinated bids, geographic partitioning, hidden shared infrastructure
FT-SYNTH	Synthetic / assembled identity	Mixed real/fake attributes, recent-only history
FT-RECON	Temporal reconstitution	Same infrastructure after debarment with variable time gap

Each typology is planted at three difficulty tiers: easy (1-hop), medium (2-hop), and hard (3+ hop structural fingerprint).

Training Data

Trained on fwa_graph_v1.pt — a synthetic heterogeneous PyG HeteroData object exported from Salesforce Data Cloud. Ground truth labels come from the Phase 0 synthetic generator (dataset/ground_truth/manifest.json).

Split	Total	Fraud	Fraud %
Train	12,013	3,087	25.7%
Val	2,559	647	25.3%
Test	2,563	649	25.3%

Note: The training set uses cluster-stratified splits to prevent data leakage — entities belonging to the same fraud ring land in the same split.

Evaluation Results

Best checkpoint: epoch 132 (lowest val loss = 0.2057).

Metric	Val Set
Val Loss	0.2057
Precision	0.797
Recall	0.983
F1	0.880
AUROC	0.9908
Average Precision	0.9721

Training converged with AUROC ≥ 0.99 by epoch ~50, with F1 continuing to improve through epoch 132 via precision gains as recall stabilized near 0.98.

Usage

import torch
from model import RGCN_FWA, build_model

# Load the graph
data = torch.load("fwa_graph_v1.pt", weights_only=False)

# Load model weights
model = build_model(data, hidden_dim=128, num_layers=2, num_bases=30, dropout=0.3)
model.load_state_dict(torch.load("model_best.pt", weights_only=True))
model.eval()

# Score Account nodes
with torch.no_grad():
    out = model(data.x_dict, data.edge_index_dict)
    logits = out["Account"].squeeze(-1)          # shape: [num_accounts]
    probs  = torch.sigmoid(logits)               # fraud probability per account

# Default threshold: 0.5
predictions = (probs > 0.5).long()

Feature Normalization

Account features must be z-score normalized before inference using the statistics in feature_stats.json:

import json, torch

with open("feature_stats.json") as f:
    stats = json.load(f)

mean = torch.tensor(stats["mean"])
std  = torch.tensor(stats["std"])

# data["Account"].x is raw; normalize before passing to the model
data["Account"].x = (data["Account"].x - mean) / (std + 1e-8)

FWA Synthetic Dataset v1

Synthetic Dataset on Hugging Face

Dataset Description FWA Synthetic Graph v1 is a synthetic knowledge graph representing federal grant and loan program entities with planted fraud typologies, designed for Graph Neural Network (GNN) training and Salesforce Data Cloud identity resolution demonstration.

The dataset models realistic federal assistance program operations including applicants, applications, payments, vendors, and agencies. It contains 925 carefully planted fraud clusters across 4 distinct typologies at 3 difficulty tiers, enabling supervised learning for fraud detection models and comprehensive testing of identity resolution workflows.

Files

File	Description
model_best.pt	Best checkpoint weights (state dict, epoch 132)
model.py	RGCN_FWA class and build_model() convenience constructor
feature_stats.json	Mean/std for 19 Account node features (for z-score normalization)
export_metadata.json	Graph provenance: node/edge counts, fraud label counts, split stats
requirements.txt	Python dependencies (PyTorch 2.11, PyG 2.7)

fwa_graph_v1.pt (the full graph, ~140 MB) is not included here — it is generated from the synthetic dataset and Data Cloud export pipeline. See the project README for reconstruction steps.

Intended Use

• Research and demonstration of GNN-based fraud detection on federal grant/loan graphs
• Baseline for comparing rule-based CIO signals vs. learned graph representations
• Integration with Salesforce Einstein Studio BYOM for on-demand scoring via Apex

Out-of-Scope Use

This model is trained on fully synthetic data and is not intended for production deployment against real federal program data without retraining on real, validated datasets with appropriate privacy controls and fairness audits.

Limitations

• Graph is static (no temporal edge streaming)
• Hard-tier fraud detection (3+ hop patterns) benefits from AUROC ~0.99 at the graph level but individual-entity precision remains ~0.80 — false positive rate requires analyst review before case creation
• Learnable embeddings for node types without dense features (Application, BankAccount, etc.) are not transferable — the model must be retrained if the graph topology changes significantly

Citation

@misc{kvirtue2026fwagnn,
  author       = {kvirtue},
  title        = {FWA GNN: Relational Graph Convolutional Network for Federal Fraud Detection},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/kvirtue/fwa-gnn-rgcn}}
}