Purpose: This study examines whether native transaction-level attributes can classify scam-labeled and non-scam-labeled Ethereum transactions using supervised machine learning under a controlled benchmarking framework, before graph-derived or address-history representations are introduced. Methods: A public dataset of 71,239 Ethereum transactions was analyzed using eleven numerical and temporal predictors covering value, gas-related attributes, transaction ordering, block metadata, and timestamp-derived characteristics. Logistic Regression, Random Forest, and LightGBM were evaluated under consistent partitioning and leakage-aware preprocessing, with SMOTE applied within training pipelines. Performance was assessed using accuracy, precision, recall, F1-score, and ROC-AUC. Findings: Ensemble models outperformed the linear baseline, indicating that nonlinear relationships among transaction attributes carry useful classification signals. LightGBM achieved the highest accuracy (0.9799), recall (0.9381), F1-score (0.9383), and ROC-AUC (0.9944), while Random Forest performed comparably with higher precision (0.9475). Block-related and gas-related attributes were prominent predictors, although these associations are not causal. Research Implication: The benchmark provides a transparent baseline for assessing whether graph-based, address-history, and temporal representations add predictive value beyond native transaction information, and supports future chronological and address-disjoint validation. Originality: Unlike prior studies relying on account aggregation, graph learning, or learned representations, this study isolates the predictive information contained in compact transaction-native attributes through a controlled, leakage-aware transaction-level benchmark, providing a reference point for quantifying the incremental value of richer representations in Ethereum scam classification.
Copyrights © 2026