Fraud Detection in Banking Transaction Network Using a Hybrid Deep Learning Method

Fraud in banking transactions is a major issue for the financial sector, whether carried out by individuals or as part of organized schemes. Detecting such illegal activities is inherently challenging. Although numerous models and algorithms have been developed to address this problem, the complexity and variability of fraud schemes make it difficult to identify all suspicious transactions. To address this, researchers have proposed using graph theory to analyze the relationships between transactions. Furthermore, significant insights can be gained by examining transaction interactions over a specific timeframe, particularly by analyzing the sequence of features and their patterns across successive transactions. This paper presents a hybrid approach using graph and deep learning models, considering sequential analysis to detect organized fraudulent activities. The proposed approach begins with four distinct scenarios: statistical data, network features, a combination of statistical features and network features, and finally a combination of egonet features with statistical features and network features, using deep learning models to detect organized fraud. The study uses a dataset of 345,917 transactions related to customer card activities to classify legal and illegal transactions. The results show that the fourth scenario, which includes egonet features along with statistical and network features and uses LSTM, performs acceptably compared to other models.