Graph Convolutional Network-Driven Adaptive Learning Framework for Fraud Detection in Complex Transactional Cryptonetworks

With a growing need to tackle the issue of fraudulent transactions in the financial domain, especially with the boom of cryptocurrency. This study proposes a Graph Convolutional Network (GCN)--based classification model for detecting fraudulent transactions by leveraging graph-structured data. Traditional classification methods often fail to capture complex topological dependencies within financial networks, whereas GCNs effectively learn from node connectivity and feature propagation. The dataset undergoes extensive preprocessing, including KNN imputation, power transformations, polynomial feature engineering, and variance thresholding, ensuring optimal feature selection. To address the class imbalance, a hybrid SMOTE-ADASYN resampling technique is employed. Experimental results demonstrate that the proposed GCN outperforms CatBoost, XGBoost, SVM, Logistic Regression, and Random Forest, achieving an accuracy of 98.75% and an AUC score of 0.97. Additionally, subgraph visualizations provide interpretability into model predictions. This work establishes GCNs as a superior alternative for fraud detection in transaction networks.