Researchers have developed DeepScrub, a new framework utilizing large language models (LLMs) for detecting fake-order fraud in online-to-offline services. This system integrates various risk signals into textual descriptions for LLMs, enhances domain knowledge through continued pre-training, and refines reasoning paths with expert feedback. In testing, DeepScrub achieved an 85.3% macro-F1 score, outperforming existing methods, and a smaller 8B model demonstrated superior performance to a larger 32B model due to effective domain adaptation. A live pilot showed significant improvements in precision and recall, drastically reducing manual review workload and saving substantial costs. AI
IMPACT This framework demonstrates how LLMs can be effectively adapted for specialized tasks like fraud detection, potentially improving accuracy and efficiency in risk management systems.
RANK_REASON Research paper detailing a novel LLM framework for fraud detection. [lever_c_demoted from research: ic=1 ai=1.0]
- 32B model
- 8B model
- alphaXiv
- arXiv
- CatalyzeX
- DagsHub
- DeepScrub
- fake-order fraud detection
- Gotit.pub
- Hugging Face
- LLMs
- ScienceCast
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
