The complete platform for Graph ML fraud defense
Designed from the ground up for high-throughput payment rails. Combining relational graph topology, gradient boosted decision trees, and sub-millisecond feature caching.
Multi-Head Graph Attention (GAT)
Unlike standard GCN isotropic averaging, GAT dynamically weighs transaction edges with multi-head attention. It identifies illicit hub routing even when bad actors deliberately disguise behavior with innocent decoy transactions.
Dual-Layer Nearline Serving
Live multi-hop GNN message passing is too slow for real-time payment SLAs (>50ms). Radar 3D computes graph embeddings nightly and caches them in a Redis 7 feature store, serving risk scores in under 0.85ms at checkout.
Imbalance-Aware Focal Loss
In extreme 0.25% fraud class imbalance, standard cross-entropy fails. Our Focal Loss objective ($\alpha=0.5, \gamma=2.0$) downweights easy negative gradients by up to 10,000x, forcing the network to focus on evasive mule signatures.
22-Dimensional Feature Engine
Combines PageRank centrality, K-core graph decomposition, in/out degree ratios, balance drain speed, night transaction ratios, and 24h/7d velocity surges into a unified feature tensor.
curl -X POST https://api.radar3d.io/v1/predict \
-H "Authorization: Bearer radar_sec_8f92a10" \
-H "Content-Type: application/json" \
-d '{
"account_id": "C_MULE_8841",
"balance_drain_ratio": 0.98,
"night_tx_fraction": 0.85,
"degree_ratio": 16.0,
"pagerank": 0.0085,
"tx_velocity_24h": 48
}'| Model Architecture | PR-AUC | ROC-AUC | F1-Score | Recall | Precision@100 | Precision@500 | Latency SLA |
|---|---|---|---|---|---|---|---|
| Logistic Regression (22 Feat) | 0.0715 | 0.6948 | 0.0172 | 74.85% | 1.0% | 46.2% | < 1.2 ms |
| LightGBM (Histogram Baseline) | 0.0106 | 0.6754 | 0.0188 | 93.23% | 2.0% | 51.4% | < 3.5 ms |
| XGBoost (Standard 22-Feat) | 0.0861 | 0.8725 | 0.0364 | 83.43% | 92.0% | 25.8% | < 5.8 ms |
| GNN — GCN (Isotropic Aggregation) | 0.0380 | 0.7420 | 0.0240 | 68.20% | 12.0% | 8.4% | ~ 65.0 ms |
| GNN — GraphSAGE (Inductive Sample) | 0.0490 | 0.7850 | 0.0310 | 72.40% | 18.0% | 12.6% | ~ 50.0 ms |
| GNN — GAT (Multi-Head Attention) | 0.0942 | 0.9129 | 0.0512 | 88.40% | 94.0% | 48.2% | ~ 85.0 ms |
| Hybrid GAT + XGBoost Ensemble PRODUCTION WINNER | 0.1185 | 0.9247 | 0.0684 | 89.60% | 96.0% | 54.2% | < 0.85 ms (Redis) |
| Two-Stage Cascade Consensus HIGH PRECISION | 0.1120 | 0.9180 | 0.0620 | 88.20% | 95.0% | 51.8% | < 0.78 ms (Redis) |
1. Why GAT Attention Beats Isotropic GCN
Standard GCN averages all incoming neighbors equally. In sparse payment graphs, an account connected to 50 normal users and 1 fraud hub suffers from over-smoothing. GAT learns dynamic edge attention weights $\alpha_{ij}$ to isolate suspicious money laundering paths:
2. Mathematical Imbalance: Focal Loss
With 0.13% fraud prevalence (130:1 class imbalance), standard cross-entropy is overwhelmed by easy negative gradients. Focal Loss downweights easy negative examples by up to 10,000x:
3. Dual-Layer Serving Architecture
Live GNN message passing takes >50ms over 3.28M nodes, violating payment gateway SLAs. Our architecture decouples nightly PyG batch inference into a Redis Feature Store for sub-millisecond scoring: