In recent years, graph-based methods have been exploited to identify anomalous activities by using graph structures and the underlying relational data. Recently, with the swift adoption of Graph Neural Networks (GNNs) for fraud detection, assessing node suspiciousness using neighborhood information has gained importance. However, large-scale datasets and imbalanced class distribution continue to be the key challenges in this domain, which can hinder model performance. Additionally, fraud patterns and behaviors may evolve over time, necessitating the adoption of incremental learning techniques to adapt to these changes. To tackle these challenges, we propose an approach that incorporates scalable clustering and effective weighted undersampling for the majority class, tackling the issue of data imbalance. Furthermore, to mitigate the problem of catastrophic forgetting when learning from new data, we utilize a regularization-based incremental learning framework called Elastic Weight Consolidation (EWC). This method enables the model to preserve previously learned knowledge while adjusting to new patterns.