Evaluating Dimensionality Reduction Techniques in Bitcoin Ransomware Detection: Comparative Analysis of Incremental PCA and UMAP

ABSTRACT This study investigates the classification performance of various machine learning algorithms on the Bitcoin Heist ransomware dataset, focusing on the effects of dimensionality reduction techniques. The primary objective was to evaluate the classifiers' effectiveness in distinguishing between malicious and benign transactions under three experimental scenarios: without dimensionality reduction, utilizing Incremental Principal Component Analysis (IPCA), and applying Uniform Manifold Approximation and Projection (UMAP). The methodology involved rigorous experimentation with four classifiers: K‐Nearest Neighbors (KNN), XGBoost, Decision Tree, and Multi‐Layer Perceptron (MLP). The results demonstrated that dimensionality reduction techniques, particularly UMAP, improved the performance of KNN and Decision Tree classifiers while adversely affecting the performance of XGBoost and MLP. Notably, KNN consistently outperformed the other classifiers across different scenarios, indicating its robustness in handling reduced feature spaces. This study concludes that the effectiveness of dimensionality reduction is contingent upon the specific characteristics of the classifiers employed.