Tackling Class Imbalance Problem with Adversarial Attention-based Variational Graph Autoencoder: Study in Fraud Detection

Abstract

Class Imbalance is often encountered in many classification problems for machine learning, resulting in bias towards the majority class. Various techniques have been developed to address this issue, focusing on oversampling the minority class or undersampling the majority class. This research aims to tackle the class imbalance problem with a different approach by using Adversarial Attention-based Variational Graph Autoencoder (AAVGA) introduced by Weng et al. in 2020. This approach is studied for fraud detection, which graph model of classification task are first constructed by mapping each entity as nodes and transaction between them as edges. The experiment is conducted by varying class distributions to analyze how imbalance class influences the prediction score. We obtained that our approach produced better precision and recall score, even for extremely imbalanced dataset.