Improving fraud detection with semi-supervised topic modeling and keyword integration

Area of Science:

• Information Security
• Computational Linguistics
• Auditing

Background:

• Traditional fraud detection relies on manual review, which is limited by human experience.
• Technological approaches, including natural language processing (NLP), face challenges in replicating manual fraud detection.
• Unsupervised topic modeling techniques like LDA and NMF have limitations for specific tasks such as fraud detection.

Purpose of the Study:

• To propose a semi-supervised topic modeling approach for improved fraud detection.
• To incorporate domain-specific knowledge using keywords to enhance topic discovery.
• To identify patterns related to the fraud triangle theory for more interpretable results.

Main Methods:

• Developed a semi-supervised topic modeling approach incorporating domain-specific keywords.
• Utilized keyword-driven learning to identify latent topics associated with fraud.
• Evaluated model performance using multiple training datasets and an independent test dataset.

Main Results:

• The proposed semi-supervised topic modeling approach demonstrated efficient performance in fraud detection.
• Achieved a 7% performance increase compared to previous methods.
• The model provided more consistent and interpretable results by identifying fraud-related patterns.

Conclusions:

• Semi-supervised topic modeling offers a promising strategy for proactive fraud detection.
• Incorporating domain knowledge through keywords significantly enhances topic modeling for fraud analysis.
• Further research into fraud behavior analysis and proactive identification strategies is crucial.