Feature Engineering for Anomaly Detection and Classification of Blockchain Transactions

Abstract

The study analysed the importance of blockchain transaction features to identify suspicious activities. The feature engineering process involves exploiting domain knowledge, applying intuition, and performing a time-consuming series of trial-and-error extractions. Manually overseeing this process significantly impacts the performance of model generation. We address this challenge with an automated feature engineering approach to extract the various features from blockchain transactions. Also, we engineered a set of new features based on statistical measures and graph representation. We demonstrate that the proposed approach can be applied to various blockchain transaction datasets, including Bitcoin and Ethereum. The engineered features were tested against eight classifiers, including random forest, XG-boost, Silas, and neural network-based classifiers to identify the suspicious behaviour of transactions