Design and development of an LLM-based framework for crime classification and prediction

Abstract

Large language models (LLMs), as a subset of generative AI, have been used in different domains, such as financial, medical, legal, and agricultural applications. However, adopting LLMs for smart policing applications remains unexplored. This thesis concentrates on developing a framework based on the transformative potential of the BART, BERT, and GPT models in this domain using methods such as zero-shot prompting, few-shot prompting, and fine-tuning. As a prototype, these methods were used to comprehensively assess the performance of LLMs in crime classification and prediction based on state-of-the-art datasets from two major cities: San Francisco and Los Angeles. The main objective is to illuminate the adaptability of LLMs and their capacity to revolutionize crime analysis practices. Additionally, a comparative analysis of the aforementioned methods on the GPT model and BART with machine learning (ML) techniques is provided. The experimental results demonstrate the feasibility of integrating LLMs into smart policing systems and show that GPT models are more suitable than traditional ML models for crime prediction in most experimental scenarios.