Metaheuristic‐Driven Feature Selection for Blockchain Fraud Detection: A Comparative Study on Machine and Deep Learning Models

Blockchain ecosystems such as Bitcoin and Ethereum are increasingly exposed to fraudulent and anomalous activities, creating a need for reliable detection methods under severe class imbalance. Although recent studies have applied machine learning (ML) and deep learning (DL) models, many rely on accuracy‐oriented evaluation and do not systematically examine feature selection, stability, and decision thresholding. In this work, we study metaheuristic feature selection using Aquila Optimiser (AO), Fox Optimiser (FOX) and Artificial Bee Colony (ABC) to improve the performance–parsimony trade‐off in blockchain fraud detection. Using public Bitcoin and Ethereum datasets, we benchmark a diverse set of models (e.g., RF, XGB, LightGBM, FTTransformer and SAINT) with imbalance‐aware metrics (PR–AUC and Precision@G1), alongside ROC–AUC for completeness. We further compare threshold‐selection strategies on a validation set and assess feature‐subset stability across multiple runs. Our results show that ABC‐selected subsets substantially reduce the number of features and training time while maintaining–often improving–predictive performance relative to full feature sets. Overall, metaheuristic‐driven feature reduction yields compact and stable feature subsets that support consistent and efficient blockchain fraud detection across ML and tabular DL pipelines.