# AI/ML Classification for Suspicious Mule Account Detection
> Learn about graph-augmented and temporally-aware ensemble classification architectures for real-time detection of financial mule accounts.

Tags: ai, machine-learning, fintech, fraud-detection, graph-networks, banking, cybersecurity
## AI/ML Classification of Suspicious Mule Accounts
- Focuses on real-time mule account detection using graph-augmented and temporally-aware ensemble classification.
- Identifies multi-tiered architectures (Collectors, Relays, Consolidators) designed to evade rule-based monitors.

## Why Standard ML Fails on Tabular Data
- Tabular features suffer from 'feature blindness' at an account level.
- Fraudsters use transaction structuring to mimic legitimate behavior.
- Graph context reveals relations (e.g., Tier 2 Relays) that tabular snapshots miss.

## The Class Imbalance Trap
- Target prevalence is typically 1-3%.
- Accuracy is misleading (98% accuracy can catch zero fraud).
- System optimizes for F1-Score and Recall: Phoenix Ensemble achieves 0.82 F1-Score compared to 0.60 baseline.

## Dual-Pipeline Stacking Ensemble Architecture
- **Pipeline A**: XGBoost on a 31-feature matrix containing tabular and temporal data.
- **Pipeline B**: 2-Layer Graph Attention Network (GAT) for structural relationships.
- **Meta-Learner**: Synthesizes a final risk score using SHAP for explainable alerts.

## Feature Engineering & Imbalance Mitigation
- Engineered signals include betweenness centrality, scatter ratio, and dormancy spike ratios.
- Four-layer correction strategy: Data (GraphSMOTE), Algorithm (Cost-sensitive weights), Loss (Focal Loss), and Threshold (PR-tuned boundary).

## Explainable AI (XAI) for Compliance
- Converts SHAP feature contributions into human-readable compliance reason codes.
- Aids in filing Suspicious Activity Reports (SAR) with audit-traceable language.

## Operational Roadmap & Benchmarks
- Near-real-time tabular scoring (<200ms latency).
- Daily batch graph recomputation.
- Performance uplift: Baseline (0.60) -> +Graph (0.72) -> +Temporal (0.76) -> Full Ensemble (0.82).
---
This presentation was created with [Bobr AI](https://bobr.ai) — an AI presentation generator.