In an evaluation on a public 9,841-record Ethereum dataset, a cost-sensitive, penalty-based HHO-XGBoost framework attained a mean false-negative rate of 0.001835 across ten held-out test sets. The framework incorporates Harris Hawks Optimization to configure classification models under an asymmetric loss structure, meeting all four pre-selected empirical feasibility targets across every test repetition.

The authors framed Ethereum fraud screening as an operating-point decision problem where missed fraudulent addresses and false alerts impose differing operational consequences. To address these consequences, the optimization objective incorporates an assumed 10:1 false-negative-to-false-positive cost ratio. The framework also applies an additive penalty whenever candidate models depart from four target feasibility metrics: Precision, false-positive rate, area under the precision-recall curve, and Matthews correlation coefficient.

To locate effective configurations, the framework conducts a joint search across an 11-dimensional space. Within this space, Harris Hawks Optimization tunes XGBoost hyperparameters, class-weight scaling, and a validation-selected decision threshold simultaneously. This joint formulation avoids optimizing parameters in isolation, ensuring that decision thresholds and tree parameters adapt together to the asymmetric cost objective and feasibility constraints.

The experimental evaluation used training-only preprocessing on the 9,841-record dataset to prevent data leakage. Researchers divided the records using a five-way group-aware partition into training, search-validation, calibration, threshold-selection, and held-out test sets. This partitioning structure was executed across ten repeated seeds, strictly maintaining zero group overlap across partitions throughout testing.

The benchmark evaluation examined five alternative optimization techniques within the identical 11-dimensional search space: Tree-structured Parzen Estimator, random search, particle swarm optimization, genetic algorithm, and Grey Wolf Optimizer. Each optimizer operated under an identical budget of 1,800 candidate evaluations per run. Across all ten held-out tests, these alternative methods recorded mean false-negative rates spanning from 0.012538 to 0.022324, exceeding the rate achieved by HHO.

Across the ten test evaluations, the HHO-XGBoost model registered a mean false-positive rate of 0.002349, Precision of 0.991802, Matthews correlation coefficient of 0.993545, and area under the precision-recall curve of 0.998957. Attribution controls examining an XGBoost model with threshold selection alone recorded a mean false-negative rate of 0.021713, confirming that threshold selection by itself cannot reproduce the full framework’s performance.

The study outlined clear constraints regarding the interpretation of these classification findings. Because the public dataset summarizes each address’s observed history and contains no timestamps, the results strictly support retrospective account-level classification within the evaluated dataset. The authors concluded that prospective deployment cannot rely on historical summaries alone, emphasizing that temporal validation and external validation are required before prospective use.

Subject of Research: Cost-sensitive machine learning optimization for retrospective Ethereum fraud detection

Article Title: Cost-Sensitive Constrained HHO-XGBoost for Ethereum Fraud Detection

Article References: Sedighimanesh, M., Hasan, S. H., Keihani, H., Barzegar, F., & Sadeghi, M. (2026). Cost-Sensitive Constrained HHO-XGBoost for Ethereum Fraud Detection. International Journal of Computational Intelligence Systems. https://doi.org/10.1007/s44196-026-01701-5

Image Credits: AI Generated

DOI: 10.1007/s44196-026-01701-5

Keywords: Ethereum fraud detection, HHO-XGBoost, Harris Hawks Optimization, cost-sensitive learning, hyperparameter optimization, false-negative rate, false-positive rate, Matthews correlation coefficient, precision-recall curve, Tree-structured Parzen Estimator, particle swarm optimization, genetic algorithm, Grey Wolf Optimizer, decision threshold selection, class-weight scaling, group-aware partition, account-level classification, retrospective validation

News Source: Denise Maddox. (November 17, 2026). Cost-Sensitive HHO-XGBoost Cuts False Negatives in Retrospective Ethereum Fraud Screening. Scienmag.