| Abstract [eng] |
The rapid growth of the financial technology (Fintech) sector has significantly increased the volume and complexity of financial transactions, creating new challenges for anti-money laundering and transaction monitoring systems. Financial are required to comply with strict regulatory frameworks while maintaining efficiency and customer-friendly payment processes. Traditional rule-based systems generate a high number of false positive alerts, leading to excessive operational costs, inefficient use of human resources, and reduced overall effectiveness of AML processes. Therefore, the application of advanced analytical methods, such as process mining and large language models, has become increasingly relevant for improving transaction monitoring efficiency. The object of paper – the efficiency of transaction monitoring process in financial institutions, with a focus on false positive alerts reduction. Paper‘s objective – to propose a method that integrates process mining and large language models in order to improve the efficiency on transaction monitoring processes by reducing the number of false positive alerts. Paper tasks: 1. To analyze the principles and limitations of traditional transaction monitoring systems in the context of anti-money laundering. 2. To examine the theoretical and practical applications of process mining techniques and large language models in financial crime detection. 3. To develop a hybrid method combining process mining and large language models for extracting additional risk-related features from transaction data. 4. To conduct an experimental evaluation of the proposed method using transaction monitoring dataset and assess its effectiveness with machine learning models. Conclusions. The results of the research demonstrate that the integration of process mining and large language models can significantly enhance transaction monitoring systems. Using process mining for identification of complex behavioral patterns and large language models for incorporation of unstructured and rarely used data contributes in generating new features that can more effectively indicate fraudulent transactions. The proposed hybrid approach contributes to a reduction in false positive alerts and improves the overall efficiency and accuracy of anti-money laundering processes. Scope of paper. The paper consists of three main parts and comprises 102 pages, including theoretical analysis, method proposal and development, and experimental evaluation. The study includes 11 tables, 14 figures, and 8 appendices. A synthesized transaction data set was used due to data protection constraints, and the experimental analysis was conducted using process mining tools, selected large language model and machine learning models for final evaluation. |
