Idea
A scalable graph-based fraud detection model that improves account takeover detection and reduces user friction for banks.
Research Paper
Core Innovation
This paper introduces ATLAS, a framework that models account takeover detection as spatio-temporal node classification on a directed session graph. It uniquely incorporates time-respecting message passing and label propagation constrained by recency and time windows, enabling causal and leakage-free learning. This approach outperforms traditional independent session scoring by capturing relational and temporal attack patterns at scale.
Market Size (TAM)
$20–50B TAM for Fraud Detection Software; $2–10B SAM from Consumer Banking and Financial Services. Driven by increasing digital fraud and regulatory compliance demands.
Potential Customers & Pain Points
- Consumer Banks Needing High Recall Fraud Detection
- Financial Institutions Reducing Customer Friction
- Fraud Prevention Teams Handling Coordinated Attacks
Business Model
Enterprise software licensing and SaaS subscription targeting financial institutions with fraud detection needs.
Competitive Landscape
- SAS Fraud Management
- FICO Falcon Fraud Manager
- Featurespace
Implementation Challenges
- Integration with existing banking infrastructure
- Data privacy and regulatory compliance
- Scalability to real-time high-volume transactions
Validation Strategy
- Pilot deployment with partner banks to measure fraud detection improvement
- Benchmark against existing fraud detection models on real transaction data
- Iterate model based on latency and user friction feedback
Research Paper Overview
Spatio-Temporal Directed Graph Learning for Account Takeover Fraud Detection
Summary
Account Takeover fraud detection is improved by modeling online sessions as a spatio-temporal directed graph linking entities via shared identifiers with time constraints. The ATLAS framework uses inductive GraphSAGE variants for scalable, causal message passing and label propagation, achieving better fraud capture and reduced user friction in production at Capital One.