Abstract: The rapid expansion of the Ethereum network has created serious security threats, which require robust machine-learning-based fraud detection systems. However, these classifiers often struggle with high dimensionality, class imbalance and computational overhead. In this study, we address these problems by a comprehensive evaluation of 10 algorithms for feature selection based on mutual information (MI) based feature selection methods combined with three tree-based ensemble classifiers: Random Forest, XGBoost and LightGBM. We utilized two real-world datasets, the Ether Fraud Detection Dataset (EFDD) and the Labelled Ethereum Transaction Dataset (LETD). This research aims to identify the most optimal combinations to maximize predictive accuracy while minimizing the number of sub-sets of the features. It is observed from empirical results that MI-based feature selection techniques successfully isolate critical transactional patterns and gradient decedent models are superior to conventional methods. Specifically, Kurtosis Based Feature Selection (KBFS) paired with LGBM achieved high Matthews correlation coefficient (MCC) of 0.9641 for EFDD using only 15 features. In LETD, using only 6 features ranked by the Double-Input Symmetric Ratio (DISR) with LGBM yielded 0.9689 MCC. Finally, the optimal selection of MI-based features reduced the computation time while keeping the most up-to-date Ethereum fraud detection.

Key words: Ethereum fraud detection, feature selection, LightGBM, mutual information, Mathew correlation coefficient, random forest, XGBoost