Enhancing Fraud Detection
summary
Company profiles are auto-generated per section using an SOP as the prompt and take about a minute or two to produce, then encoded as embeddings or tokenized n‑grams. A simple pipeline of TF‑IDF n‑grams + logistic regression surfaces which sections carry signal (transactions and top counterparties lead), and embeddings from LLM‑generated sectioned summaries outperform raw text and many manual features, with one discovered feature giving a ~20% relative recall lift; caveats: hallucination, non‑determinism, and latency.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 1:49Fraud types and challenges
- 3:59Feature engineering
- 7:33Automating reviews with LLMs
- 13:57Signal from summaries
- 18:35Discovered features and improvements
- 22:32More use cases
- 25:21Production challenges
- 27:08Summary
related talks
28mFortifying Digital Trust: Next-Gen Strategies in Fraud Prevention
35mGenerative AI for Fraud Detection & Claims Analysis
52mModern Fraud Prevention Using Deep Learning
46mFraud Detection at Floa
16mAI-Driven Behavioral Analytics for Real-Time Fraud Prevention
26mAdvancing Financial Fraud Detection with Graph Dbs
08mAI-Powered Fraud Detection in Financial System
18mReal-time Fraud Detection using Deep Learning