| Abstract [eng] |
This thesis describes research about machine learning models usage for suspicious transactions detection. The research was done using IBM corporation ‘s synthetic database generated with AMLworld simulator, which resembles real money laundering scenarios. The goal of research is to compare unsupervised machine learning models in suspicious transaction detection. In the research, static time-based features are separated as well as dynamic frequency-based transactions and group features by dividing users to groups by their payment averages. Features are chosen by importance and correlation with money laundering label from dataset. Chosen features are preprocessed using PCA dimensionality reduction technique, data is balanced using SMOTE algorithm. Processed features are classified using gradient boosting, logistic regression classifiers. An alternative approach to the research is also investigated, assuming that legal transactions are similar and with similar features they can be assembled to one or several clusters. Data objects are assigned to clusters using DBSCAN clustering algorithm, while using different fine-tuning parameters. Outliers are assumed to be illegal transactions in this approach. There is also used anomaly detection algorithm Isolation Forest which separate anomalies without clustering. For reduction of false positive results an ensemble of algorithms is used which gives voting right to every algorithm and the weight of the vote is determined by the F1 qualitative score. After conducting research, the conclusion is made that PCA and SMOTE algorithms give improvements on classification prediction. |
