Unsupervised deep learning to supervised interpretability : a dual-stage approach for financial anomaly detection

• Ano de defesa: 2024
• Autor(a) principal: Freire, Marcos Brum
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Pontifícia Universidade Católica do Rio Grande do Sul; Escola Politécnica; Brasil; PUCRS; Programa de Pós-Graduação em Ciência da Computação
• Palavras-chave em Português: Financial Anomaly Detection; Self-Adversarial Variational Autoencoders; Explainable Boosting Machines; Deep Learning; Interpretable AI; Detecção de Anomalias Financeiras; Autoencoders Variacionais Auto- Adversariais; Máquinas de Impulso Explicáveis; Inteligência Artificial Explicável; Aprendizado Profundo.; CIENCIA DA COMPUTACAO::TEORIA DA COMPUTACAO
• Link de acesso: https://tede2.pucrs.br/tede2/handle/tede/11662

The increasing sophistication of money laundering activities demands approaches that unite effective anomaly detection with interpretability. To address this challenge, we propose a dual-stage architecture integrating a Self-Adversarial Variational Autoencoder with transformer blocks for unsupervised anomaly detection, paired with an Explainable Boosting Machine for supervised classification. This approach addresses fundamental limitations in financial fraud detection, such as the scarcity of labeled data and extreme class imbalance. In evaluations on proprietary financial transaction data, the framework achieved a Receiver Operating Characteristic Area Under the Curve of 0.9508 and a Precision-Recall Area Under the Curve of 0.5417. When applied to the public credit card fraud dataset, the model attained a ROC AUC of 0.964, outperforming established methods in the literature such as Deep Autoencoder (0.882) and Autoencoder with Clustering (0.961), despite not using labeled data during training. The Explainable Boosting Machine component enabled clear identification of factors driving risk classifications, while the Self-Adversarial Variational Autoencoder component proved effective in detecting anomalous patterns across different financial contexts. The results demonstrate the potential of this integrated solution, which combines advanced detection capabilities with the transparency necessary for practical applications in the financial sector