Machine learning algorithms and visualization techniques for financial anomaly detection
Dehnavi, Nastaran (2025)
2025
All rights reserved. This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2025060520612
https://urn.fi/URN:NBN:fi:amk-2025060520612
Tiivistelmä
The aim of this thesis is to analyze anomalies, more specifically suspicious bank transactions, through supervised and unsupervised machine learning algorithms. The core research query is how various machine learning models detect and analyze unusual patterns occurring in transaction data and picking out the best algorithms for fraud detection. This study involves supervised learning techniques such as KNN, Decision Tree, Random Forest, SVM, and Gradient Boosting. On the other hand, unsupervised learning algorithms include K-Means, DBSCAN, Isolation Forest, and Local outlier factor (LOF).
This is a practical study based on the report on bank transactions. Various preprocessing data techniques are applied first. Then, different supervised and unsupervised models are trained and monitored through the process of fraud detection. Supervised models are monitored and compared using the metrics of accuracy, recall, and F1-score; important insights were then extracted through feature analysis. For unsupervised models, there was analysis based on the number of detected outliers and cluster distributions; visual analysis and pattern comparisons are also taken into consideration. Feature importance is analysed with Permutation Importance to detect which are the most influential factors for fraud detection.
The results indicate that Machine Learning algorithms, especially Random Forest and Gradient Boosting, are the most useful methods for the detection of fraudulent transactions. SVM can also successfully detect fraudulent transactions, provided the data set size is not very large. Unsupervised techniques such as Isolation Forest, LOF, and K-Means are also utilized for the detection of abnormal transactions. Hence, the recommendation from this analysis is that fraud detection systems in banks should use a combination of both supervised and unsupervised models to increase the accuracy of detection. For feature importance, Decision Tree, Random Forest, and Gradient Boosting are recommended.
This is a practical study based on the report on bank transactions. Various preprocessing data techniques are applied first. Then, different supervised and unsupervised models are trained and monitored through the process of fraud detection. Supervised models are monitored and compared using the metrics of accuracy, recall, and F1-score; important insights were then extracted through feature analysis. For unsupervised models, there was analysis based on the number of detected outliers and cluster distributions; visual analysis and pattern comparisons are also taken into consideration. Feature importance is analysed with Permutation Importance to detect which are the most influential factors for fraud detection.
The results indicate that Machine Learning algorithms, especially Random Forest and Gradient Boosting, are the most useful methods for the detection of fraudulent transactions. SVM can also successfully detect fraudulent transactions, provided the data set size is not very large. Unsupervised techniques such as Isolation Forest, LOF, and K-Means are also utilized for the detection of abnormal transactions. Hence, the recommendation from this analysis is that fraud detection systems in banks should use a combination of both supervised and unsupervised models to increase the accuracy of detection. For feature importance, Decision Tree, Random Forest, and Gradient Boosting are recommended.