Answers are grounded in this paper and cite their sources.
Ethereum fraud screening is an operating-point decision problem in which missed fraudulent addresses and false alerts impose different operational consequences. This study develops a cost-sensitive, penalty-based HHO-XGBoost framework that jointly searches XGBoost hyperparameters, class-weight scaling, and a validation-selected decision threshold while penalizing departures from four empirical feasibility targets: Precision, false-positive rate (FPR), area under the precision-recall curve (AUPRC), and Matthews correlation coefficient (MCC). The evaluation protocol uses a public 9,841-record Ethereum dataset, training-only preprocessing, a five-way group-aware partition into training, search-validation, calibration, threshold-selection, and held-out test sets, and ten repeated seeds with zero group overlap across partitions. Harris Hawks Optimization (HHO), Tree-structured Parzen Estimator (TPE), random search, particle swarm optimization (PSO), genetic algorithm (GA), and Grey Wolf Optimizer (GWO) are compared in the same 11-dimensional search space under an identical candidate-evaluation budget of 1,800 evaluations per optimizer run. The objective applies an assumed 10:1 false-negative-to-false-positive cost ratio plus an additive penalty for feasibility-target violations. Across ten held-out tests, HHO achieved a mean FNR of 0.001835, FPR of 0.002349, Precision of 0.991802, MCC of 0.993545, and AUPRC of 0.998957, meeting all four empirical targets in 10/10 repetitions. Its mean FNR was lower than the equal-evaluation alternatives (0.012538–0.022324) and the threshold-only XGBoost control (0.021713). The attribution controls show that threshold selection alone does not reproduce the full result. Because the dataset summarizes each address’s observed history and contains no timestamps, these findings support retrospective account-level classification within the evaluated dataset; temporal and external validation are required before prospective use.
No takes yet. Share an insight, caveat, or question.
Sedighimanesh et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: