Leveraging Few-Shot Learning in Generative AI for UPI Transaction Classification

Classifying financial transactions based on individual needs is a vital and often complex task. Ranging from a college student who manages monthly expenses based on their allowance to a working professional preparing an annual income tax report, both require a tailored classification of finances to ensure smart spending and financial management. The recent invention of Unified Payments Interface (UPI) and constantly increasing UPI IDs have made classifying finances an extremely tedious and complex task, making supervised learning an impractical option. To tackle this problem, we propose a novel Few-Shot Learning architecture leveraging state-of-the-art Large Language Models (LLM), offering high accuracy in multi-class data classification. This prompt-based 3-way 4-shot learning utilizes the Llama 3.1 8 billion parameters model, which provides fast and efficient results on compact systems while ensuring privacy. Key findings include the importance of prompt engineering while highlighting the efficiency of few-shot learning in the finance sector. It also highlights the efficiency of open-source LLMs in task-focused applications without relying on resource-intensive models.