INTERMEDIATE // MULTI LAYER STACKING
MODULE 06 // AGGREGATION
Score Fusion.
Combine scores from multiple layers into unified risk assessments.
FUSION STRATEGIES
Each layer returns its own risk score. Score fusion combines these into a unified assessment. Simple approaches include maximum (most conservative), average, or weighted average. Sophisticated approaches use ML-based fusion.
Weighted fusion requires calibration: what weight should each layer receive? Start with equal weights, then adjust based on observed predictive power. Layers that better distinguish fraud from legitimate users receive higher weights.
Handle missing scores gracefully. If device intelligence times out, fusion shouldn't fail. Use default scores or adjust weights to exclude missing layers. Document fusion behavior when layers are unavailable.
MAX
Maximum
Use highest score from any layer. Most conservative, may increase false positives.
AVG
Average
Simple average across layers. Balanced approach, requires layer calibration.
ML
ML Fusion
Trained model combines layer signals. Best accuracy, requires labeled training data.
KNOWLEDGE CHECK // Q06
Why might weighted average outperform simple maximum fusion?