The object of this study is the process of classifying illicit Bitcoin transactions in blockchain datasets. The problem addressed in this work is the difficulty of detecting suspicious activity in cryptocurrency networks due to the high dimensionality of transaction data and the lack of semantic labels, which limits the effectiveness of conventional manual feature engineering. The proposed method combines domain-specific indicators of illicit behavior with a Genetic Algorithm-driven selection mechanism that dynamically evolves informative feature subsets. The developed framework was implemented and evaluated on the Elliptic and Elliptic++ datasets using random forest. The results obtained demonstrate that the GA-based method significantly increases model performance: the best-performing configuration achieved an F1-score of 84.3%, a precision of 99.4%, and a recall of 73.1%. Compared to baseline approaches on the same dataset, this method provides relative improvements of 0.9% in F1-score, 0.3% in precision, and 1.2% in recall. The effectiveness of the proposed solution is explained by its ability to detect hidden patterns in transactional data with many potential attributes without resorting to manual heuristics, as well as an optimized setting of Genetic Algorithm parameters. A distinctive feature of this method is the combination of heuristic search with domain-informed feature categories, which improves classification accuracy and reduces model complexity. The obtained results can be applied in practical scenarios such forensic analysis of cryptocurrency transactions. However, successful implementation requires access to historical transaction records and sufficient computing resources to process large, feature-rich datasets