Transaction Analysis of Categorizing Ethereum Addresses Based on Advanced Supervised Machine Learning Approach for Predictive Modeling

Abstract

Ethereum has become one of the most significant cryptocurrencies in terms of transaction volume. Given Ethereum's recent rise, experts and the cryptocurrency community are eager to learn more about how Ethereum transactions behave. A machine learning system-based methodology exists for addressing Ethereum addresses to enable transaction classification. The preprocessing of CEAT dataset containing 4,371 entries with 15 features utilizes a systematic process that selects relevant features then handles missing data along with using SMOTE for class distribution balancing and converting categorical elements to numbers. The data is standardized with MinMax Scaler to improve model performance. Exploratory data analysis is done by visualizations such as Heat map, Histogram and Pair plot to compute the features which are correlated. The four models of machine learning algorithms include Decision Tree (DT), LightGBM, KNeighbors Classifier (KNN), and CatBoost Classifier, are trained with the best optimized hyperparameters. The classification reports, confusion matrices, and ROC curves are used for model evaluation for each model. Comparing these models, LightGBM has the highest accuracy of 91.99%, second is CatBoost Classifier with 91.23%, the Decision Tree is 82.43%, and the KNN model 78.91%. An important benefit from this study is that the results show that it is possible to use a machine learning approach for classifying Ethereum addresses to enhance transaction security and avoid fraud in decentralized financial systems.