VynFi/je-fraud-gnn

VynFi JE Fraud GNN — v5.29 retrain

GraphSAGE edge classifier trained on the v5.29 SOTA-mode VynFi/vynfi-journal-entries-1m accounting-network edges. Binary fraud detection at the edge level (single journal entry's debit→credit relation).

Metrics on test split (held-out)

metric	v5.27 baseline	v5.29 retrain	Δ
AUC-ROC	0.909	0.9185	+0.010
AUC-PR	0.799	0.8048	+0.006
F1	0.790	0.8010	+0.011

Trained 50 epochs on NVIDIA A10, ~2.5 min. n_test = 234,953 edges, 13,776 positive (5.86 % fraud rate).

Per-process breakdown (test split)

process	n	n_pos	AUC	PR	F1
P2P	70,635	4,230	0.914	0.800	0.798
O2C	82,521	4,785	0.919	0.803	0.801
R2R	46,758	2,727	0.925	0.812	0.803
H2R	23,318	1,315	0.920	0.810	0.808
A2R	11,721	719	0.915	0.807	0.802

Repro

# Build PyG dataset from the v5.29 HF dataset
python3 scripts/ml/build_je_pyg_dataset.py \
    --output je_pyg_v2.pt --seed 42

# Train
python3 scripts/ml/train_je_fraud_gnn.py \
    --dataset je_pyg_v2.pt \
    --output je_fraud_gnn.pt \
    --epochs 50 --device cuda --seed 42

Source: mivertowski/SyntheticData @ v5.29.0.

The PR-AUC and F1 lifts over v5.27 reflect the SOTA-N behavioral levers (recurring archetypes, source-conditional rarity tagging, trading-partner pool concentration, etc.) producing sharper fraud-vs-clean separability in the accounting-network substrate.