FraudLens

From Alerts to Insights: Responsible AI for Fraud Detection

This website showcases FraudLens, the Graduate Project developed for the MS in Business Analytics program at The George Washington University School of Business. It highlights our research, methodology, and applied analytics work.

Jump to the Demo

Executive Summary

Fraud Analysts are working every day combatting fraud, but every day the task gets larger and more complex.

How are we able to help our analysts be more efficient in their efforts?

PROBLEM

Evidence is scattered, making investigations cumbersome.
Current explanations are rigid and rules-based, lacking context.

SOLUTION

Aggregate key evidence, patterns, and similar past cases.
Apply Explainable Boosting Model (EBM) + Shapley Additive Explanations (SHAP) to surface clear, defensible risk drivers.
Deliver concise, data-anchored rationales for every decision.

CONTEXT

Detecting possible fraud is mature; post-flag review is a bottleneck.
Analysts manually gather data- a slow, tedious, and leads to alert fatigue.

EXPECTED

30–50% faster fraud investigations.
$1M in labor savings per 100M transactions.
Auditable, transparent AI outputs for compliance.

FOREGROUND

Mature rule-based and machine learning (ML) models already flag suspicious transactions.

Detection is solved but explanation and reasoning remain manual.

TRIGGER

Alert volumes rising.

More cases, same staff.
- Review bottleneck and alert fatigue.

Increasing pressure for faster, auditable decisions.

IMPACT

Slow, tedious post-flag reviews.

Low recovery rate due to delayed response.

Hard-to-defend rationales in compliance audits.

STAKEHOLDERS

Capital One:
- Efficiency focus.

EY:
- Bias control & auditability.

FI Consulting:
- Validation & responsible AI.

Mizuho Financial Group:
- Explainability & analyst usability.

Fraud Analysts Surveyed

What is Your Biggest Bottleneck in Fraud Alert Processing?

Gathering supporting information

67%

Where Do You Feel AI Could Save You Time?

Summarize Alerts & Provide Explanation

44%

1. Data Preparation

• Merged sources, cleaned data, and removed personal identifiers.
• Created behavioral features and consolidated merchant categories

Architecture Diagram

Explainability

Local Case-Level Explanation

What FraudLens displays when an analyst queries a transaction.

Chatbot Explainability Output

• Highlights top contributing features via ranked logit impacts and a SHAP/EBM waterfall plot.

• Shows the model’s decision and EBM-predicted fraud probability.

• Provides a concise, RAG-enhanced narrative to translate model outputs into analyst-ready insight.

Includes traceable sources to support transparency and auditability.

• Each bar quantifies how a specific feature pushed the fraud score higher or lower for this transaction.
• EBM + SHAP deliver defensible, case-level transparency by revealing the exact factors driving the model’s decision.

Quality Impacts

Faster Investigation Turnaround

Automated retrieval and evidence generation reduces review time by up to 30-50%.

Improved Consistency of Outputs

Template-driven narrative summaries ensure uniformity across post-flag cases.

Enhanced Analyst Trust

Transparent model reasoning supports explainable, defensible case documentation.

Monetization

Demo

Risks & Mitigations

Data privacy

Implements encryption and role-based access, ensuring no sensitive data exposure.

Model Bias

Initial model explains decisions transparently; formal AIR-based fairness testing will be introduced to evaluate demographic impact.

Overreliance on AI

Empowers analysts with contextual insights instead of automated verdicts.

Auditability

Ensures every decision path is reproducible for compliance and governance.

Next Steps

1. Expand behavioral profiling:
Add customer-level behavioral profiles for deeper contextual accuracy.

2. Add AIR-based fairness testing:
Introduce Adverse Impact Ratio (AIR) to evaluate demographic fairness and ensure responsible model behavior.

3. Strengthen RAG grounding:
Tighten evidence retrieval to reduce hallucinations.

4. Enhance behavioral anomaly detection: Combine sequential features (merchant chains, spending trajectories) with network-level relationships to expose complex, realistic fraud patterns.