Fraud Analytics with Graph Databases
Staying ahead of ever-evolving fraud risks to proactively identify and investigate active threats can be a challenge for any organization. Healthcare, insurance, and financial services are just a few industry sectors that require robust fraud detection solutions. Typically, only a small fraction of fraudulent transactions are detected, and far less are prosecuted or recovered. Comprehensive efforts in fraud analytics must consider how to use data and network analysis to detect fraud rings and organized crime, whether they relate to kickback schemes, money laundering, identity theft, or other coordinated crime. In one example, a coordinated fraud ring of at least 25 members, including doctors and lawyers, filed more than $400 million in fake injury claims from staged car crashes.
Having limited samples of known fraud cases presents a significant challenge for analysts to model fraud effectively. Selection bias is profound, as the majority of claims have no characteristics caught by basic rules, leaving no labeled cases for these "passed" cases. To overcome these challenges, fraud analytics uses anomaly detection to reveal missed and emerging types of fraud.
The available feature space for real-time fraud detection is inherently small, as perpetrators skillfully minimize the evidence they leave behind. Fortunately, network analysis techniques such as graph database technologies effectively model relationships between known and suspected perpetrators. Experience has shown that these network features add enormous power to models using only traditional structured data.
This course brings these concepts together in a consistent and powerful framework to detect fraud rings. Students work through the systematic bias challenges, create predictive network features with graph databases, build powerful predictive models with the labeled data, and add additional layers to the predictive model with unsupervised techniques.
Participants will learn about:
- Data essentials: Claims, investigations, outcomes
- Bias challenges: Existing business filters
- Bias mitigation techniques
- Anomaly detection with CADE
- Combining supervised and unsupervised fraud modeling techniques
- Fraud feature engineering
- Entity Resolution
- Graph tool/technology review
- Cypher for graphs
- Network visualization
- Network feature extraction for fraud modeling
- Score explanations for fraud investigators: “Why did model ask for this claim to be investigated?
Intended Audience: Data Scientists, Data Analysts, Business Intelligence Analysts, Analytics Project Managers, or anyone interested in using graph databases for fraud detection
Duration: 3 Days