Entity Embeddings Excel in High-Cardinality Fraud Detection

When tackling fraud detection with datasets containing many unique categorical values (high-cardinality features), the choice of encoding method significantly impacts model performance. A recent study evaluated seven different categorical encoding techniques on a large fraud benchmark dataset. The goal was to compare both interpretable and learned encoding approaches. The results showed that entity embeddings, a learned encoding method, delivered the highest AUC-ROC score, demonstrating a statistically significant tie with CatBoost, which uses its own integrated encoding. Simpler methods like target encoding performed slightly worse than tier group encoding, which aims for more auditor-friendly boundaries. Notably, TabNet, a neural network approach, did not outperform tree-based pipelines and struggled with data scarcity. While entity embeddings led on AUC-ROC, CatBoost showed superior performance on AUC-PR, indicating that no single encoder dominated both metrics. The analysis confirmed that the advantage of embeddings stems from their ability to create joint, multi-column representations.