6
Models Compared
3 Deep / 2 Tree / 1 Neural
5
Scenarios Loaded
3 Easy / 2 Moderate
CAVE
Best Overall
Avg. anomaly score 0.834
CID003
Hardest Scenario
Avg. detection 0.663

Detection Heatmap — Models × Scenarios

Click cell → Explain
Model CID001
Romance Fraud
CID002
Drug Trafficking
CID003
Human Trafficking
CID004
Underground Banking
CID005
Cannabis
AVG
CWAE
0.92#2Explain →
0.88#3Explain →
0.71#2Explain →
0.83#2Explain →
0.79#5Explain →
0.826
Isolation Forest
0.85#5Explain →
0.91#1Explain →
0.62#5Explain →
0.74#5Explain →
0.87#1Explain →
0.798
Extended IF
0.87#4Explain →
0.89#2Explain →
0.65#4Explain →
0.77#4Explain →
0.85#3Explain →
0.806
Deep IF
0.89#3Explain →
0.86#4Explain →
0.68#3Explain →
0.80#3Explain →
0.82#4Explain →
0.810
AE
0.83#6Explain →
0.84#5Explain →
0.59#6Explain →
0.76#6Explain →
0.81#6Explain →
0.766
CAVE
0.94#1Explain →
0.87#4Explain →
0.73#1Explain →
0.85#1Explain →
0.78#6Explain →
0.834
Reading the heatmap: Each cell shows the anomaly score (0-1) a model assigns to a scenario, with rank among all models. Darker red = higher anomaly score = model considers the scenario more suspicious. Click any cell to drill into the Explainability view for that model × scenario combination.

Typology Detection Profile — Radar Comparison

All Models
Placement Layering Integration Structuring Network Behavioral
CWAE
Isolation Forest
CAVE
AE (dashed)
Extended IF & Deep IF omitted
for readability (similar to IF)
Insight: CWAE and CAVE (deep generative models) show broad, balanced coverage across all typologies, with particular strength in Behavioral and Network anomaly detection. Isolation Forest excels at Structuring and Placement (tabular point anomalies) but underperforms on network and layering patterns. AE is a moderate generalist with no dominant strength.

Channel Capture Analysis

Bias Check
CWAE
22%
20%
18%
15%
Isolation Forest
38%
24%
14%
Extended IF
32%
25%
15%
Deep IF
28%
26%
16%
AE
30%
25%
15%
CAVE
20%
19%
18%
16%
ABM
Cash
Cheque
EFT
EMT
Wire
Connex
⚠ Isolation Forest derives 38% of anomaly signal from Cash channel alone — potential over-capture bias toward structured cash deposits.

Graph vs Tabular Detection

Anomaly Domain
CWAE
65%
Graph 65% Tab 35%
Isolation Forest
15%
Graph 15% Tab 85%
Extended IF
20%
Graph 20% Tab 80%
Deep IF
35%
Graph 35% Tab 65%
AE
25%
Graph 25% Tab 75%
CAVE
70%
Graph 70% Tab 30%
Key takeaway: Tree-based models (IF, Extended IF) are almost entirely tabular — they detect point anomalies in feature space but miss network/relational patterns. CWAE and CAVE leverage graph structure, making them better at detecting layering and money mule networks. Deep IF sits in between as a hybrid.

Anomaly Type Distribution

Type A vs B
CWAE
30%
70%
Isolation Forest
75%
25%
Extended IF
68%
32%
Deep IF
50%
50%
AE
55%
45%
CAVE
25%
75%
Type A: Point anomalies (outlier amounts, single unusual txns)
Type B: Contextual / collective (behavioral patterns over time)
Why this matters: AML investigators care primarily about Type B (behavioral patterns like structuring, rapid in/out, escalation). Models biased toward Type A may generate excessive false positives on legitimate large transactions while missing subtle schemes.

Model Consensus

Agreement
CID001
Romance Fraud
6/6
Strong
CID002
Drug Trafficking
6/6
Strong
CID003
Human Trafficking
3/6
Weak
CID004
Underground Banking
5/6
Moderate
CID005
Cannabis
6/6
Strong
CID003 (Human Trafficking) shows weak consensus — only contextual models (CWAE, CAVE, Deep IF) detect it. This is expected: HT patterns involve subtle behavioral signals (late-night hotels, multi-city movement) that tabular models miss. CID004 is moderate: IF struggles with underground banking patterns that span cheque + wire channels.

Model Summary Cards

Quick Reference
CWAE
Contextual Wasserstein Autoencoder · Deep Generative
Strengths
  • Contextual anomaly detection via distributional matching
  • Strong on layering and network patterns
  • Balanced channel coverage (no over-capture)
Best Scenarios
  • CID001 Romance Fraud (0.92)
  • CID002 Drug Trafficking (0.88)
  • CID004 Underground Banking (0.83)
Explore Explainability →
Isolation Forest
Tree Ensemble · Tabular Isolation
Strengths
  • Fast, interpretable, low compute cost
  • Excellent at point anomalies (structured cash)
  • Best on CID002 Drug Trafficking (0.91)
Weaknesses
  • 38% signal from Cash — over-capture risk
  • Misses network-based patterns entirely
  • Worst on CID003, CID004 (contextual schemes)
Explore Explainability →
Extended IF
Extended Tree Ensemble · Improved Isolation
Strengths
  • Handles feature interactions better than IF
  • Consistent performer across scenarios
  • Lower channel bias than standard IF
Best Scenarios
  • CID002 Drug Trafficking (0.89)
  • CID001 Romance Fraud (0.87)
  • CID005 Cannabis (0.85)
Explore Explainability →
Deep IF
Neural + Tree Hybrid · Latent Isolation
Strengths
  • Combines neural representation with isolation
  • Balanced Type A / Type B detection (50/50)
  • Moderate graph awareness (35%)
Best Scenarios
  • CID001 Romance Fraud (0.89)
  • CID002 Drug Trafficking (0.86)
  • CID005 Cannabis (0.82)
Explore Explainability →
AE
Autoencoder · Reconstruction Error
Strengths
  • Simple architecture, easy to train
  • Decent generalist — no catastrophic failures
  • Good baseline model for comparison
Weaknesses
  • Lowest avg. score (0.766) — rarely best
  • Struggles most on CID003 (0.59)
  • EMT over-capture at 30%
Explore Explainability →
CAVE
Contextual Anomaly via Variational Encoding · Deep Generative
Strengths
  • Highest avg. score (0.834) — best overall
  • Best on CID001 (0.94) and CID003 (0.73)
  • Most balanced channel coverage of all models
Weaknesses
  • High compute cost for training
  • Less interpretable than tree models
  • Weakest on CID005 Cannabis (0.78)
Explore Explainability →