Uncertainty-Native Fraud Detection: Principled Multi-Source Signal Fusion with Subjective Logic
SLFD is a research framework that applies Subjective Logic (Jøsang, 2016) to fraud detection pipelines, replacing scalar fraud scores with structured opinions that explicitly represent belief, disbelief, uncertainty, and base rates.
This is not a new fraud detection algorithm. It is a data representation and fusion framework that makes existing fraud detection systems more transparent, auditable, and epistemically honest.
- Opinion construction from heterogeneous signal sources (ML models, rule engines, database lookups)
- Multi-source fusion with mathematically principled operators (cumulative, averaging, robust)
- Conflict detection that identifies disagreement between sources
- Three-way decisions (block / approve / escalate) grounded in cost-sensitive boundaries
- Temporal decay that degrades stale signals toward uncertainty
- Provenance chains for regulatory audit (BSA/AML, EU 6AMLD, PSD2)
Pre-alpha research prototype. This accompanies a research paper in preparation.
# Core only (numpy + pandas)
pip install -e .
# With ML dependencies
pip install -e ".[ml]"
# Full development environment
pip install -e ".[dev]"pytestPaper in preparation. Citation details will be added upon publication.
MIT