ENHANCING FINANCIAL SECURITY BY MODELING SEMANTIC SPENDING PERSONAS: THE STAD FRAMEWORK

Abstract

Detecting advanced credit card theft continues to be a persistent difficulty, as criminal behaviors increasingly mimic typical spending habits. Traditional systems, which mostly look at numbers like transaction amount and location, often don't respond to unusual situations. This creates a weakness where fraudulent transactions that stay within normal numerical limits go undiscovered since the rich semantic information in transaction descriptions is still not being used enough. This paper presents Semantic-Transactional Anomaly Detection (STAD), an innovative hybrid system designed to address this deficiency. The technique uses a Transformer-based language model to turn unstructured transaction stories into useful semantic vectors. A Gated Recurrent Unit (GRU) network then processes these vectors in the order they were received to create a dynamic "persona vector" for each cardholder. This vector shows how they normally shop. To find anomalies, you calculate the cosine distance between a new transaction's vector and the established persona. This gives you a semantic anomaly score. In addition to regular transactional data, this score is used as a powerful enhancement for a final XGBoost classifier. This approach enhances the model's ability to identify purchases that are numerically plausible yet semantically incongruent with a user's historical data. The STAD framework adds a more advanced, context-aware layer of security that makes it much easier to find advanced fraud cases and shows how useful deep semantic information can be for behavioral modeling.