Modeling interdependencies in fraud detection: a graph embedding-based approach

The growing incidence of fraud has made fraud detection a critical and essential focus for many companies and organizations. Fraud can often be traced based on the relationships between parties, such as fraudsters, collaborators, or victims. Understanding these relationships is essential in fraud detection, as fraud schemes typically exploit interpersonal connections and dependencies. Analyzing these interdependencies provides deeper insights into fraudulent activities, leading to more effective and higher-performing fraud detection solutions compared to methods that only consider direct relationships or lack this relational context. One of the most effective methods for examining relationships between parties is graph analysis. However, conventional graph analysis methods often produce unstructured data, making the analysis costly and complex. In this paper, we propose Graph Aware Fraud (GAF) , a solution for analyzing relationships and interdependencies in fraud detection through graph embedding. This approach represents unstructured and complex relationships in a low-dimensional feature vector using deep learning-based graph embedding techniques. The resulting feature vector can be easily used with various machine learning methods, eliminating the need for complex graph analysis algorithms. We evaluated GAF using the Healthcare Provider Fraud Detection Analysis dataset, which includes diverse parties and relationships. The experimental results revealed the model’s effectiveness, achieving a balanced F1-score of 56%. Comparative analysis with deep learning-based methods and conventional algorithms demonstrated GAF’s superiority, showcasing significant performance improvements. These results underscore GAF’s potential to uncover hidden relational patterns and enhance fraud detection. These findings highlight GAF’s potential to enhance fraud detection by uncovering hidden relational patterns in data.