Publicação

Machine Learning Algorithms applied to Credit Card Fraud Detection: A Comparative Analysis of Models Performances

Ver documento

Detalhes bibliográficos
Resumo:This dissertation addresses the growing challenge of credit card fraud detection amidst increasing online consumer transactions. The research assesses the effectiveness of various learning algorithms, comparing traditional, ensemble, and deep learning models in identifying fraudulent activities. Both empirical and theoretical approaches were employed, analyzing synthetic and real transaction data. Key performance metrics included accuracy, precision, recall, F1-score, and AUC-ROC. The study also evaluated model attributes such as interpretability and ease of modification, utilizing cross-validation and hyperparameter tuning to ensure robust results. The findings indicate that ensemble algorithms, notably Random Forest and Gradient Boosting, excel in accuracy, while deep learning models are proficient at detecting complex fraud patterns but lack interpretability. Features emphasizing temporal aspects and customer behavior significantly boost performance. Despite these advancements, challenges remain, including class imbalance, the trade-off between model complexity and interpretability, and the necessity for real-time detection. These challenges underscore the difficulty of developing a universally effective fraud detection system. In conclusion, while machine learning algorithms show significant promise for fraud detection, no single solution fits all scenarios. A combination of multiple models often produces better outcomes. This research provides practical insights for financial institutions and regulators, recommending future efforts to focus on developing real-time detection algorithms, improving model interpretability, exploring adaptive learning techniques, and investigating federated learning for diverse, privacy-preserved data use. These steps will enhance current fraud detection systems, making them more efficient, transparent, and adaptable, thereby bolstering financial security.
Autores principais:Jegbefumwen, Ehis
Assunto:Machine Learning Algorithms Credit Card Fraud Detection Real-Time Detection SDG 8 - Decent work and economic growth
Ano:2024
País:Portugal
Tipo de documento:dissertação de mestrado
Tipo de acesso:acesso aberto
Instituição associada:Universidade Nova de Lisboa
Idioma:inglês
Origem:Repositório Institucional da UNL
Descrição
Resumo:This dissertation addresses the growing challenge of credit card fraud detection amidst increasing online consumer transactions. The research assesses the effectiveness of various learning algorithms, comparing traditional, ensemble, and deep learning models in identifying fraudulent activities. Both empirical and theoretical approaches were employed, analyzing synthetic and real transaction data. Key performance metrics included accuracy, precision, recall, F1-score, and AUC-ROC. The study also evaluated model attributes such as interpretability and ease of modification, utilizing cross-validation and hyperparameter tuning to ensure robust results. The findings indicate that ensemble algorithms, notably Random Forest and Gradient Boosting, excel in accuracy, while deep learning models are proficient at detecting complex fraud patterns but lack interpretability. Features emphasizing temporal aspects and customer behavior significantly boost performance. Despite these advancements, challenges remain, including class imbalance, the trade-off between model complexity and interpretability, and the necessity for real-time detection. These challenges underscore the difficulty of developing a universally effective fraud detection system. In conclusion, while machine learning algorithms show significant promise for fraud detection, no single solution fits all scenarios. A combination of multiple models often produces better outcomes. This research provides practical insights for financial institutions and regulators, recommending future efforts to focus on developing real-time detection algorithms, improving model interpretability, exploring adaptive learning techniques, and investigating federated learning for diverse, privacy-preserved data use. These steps will enhance current fraud detection systems, making them more efficient, transparent, and adaptable, thereby bolstering financial security.