Temporal-Aware Fraud Detection Using Knowledge Graph, Embeddings and Variable Change Analysis: An Evidence Based Risk Scoring Framework

The exponential growth in digital payment systems has fundamentally transformed global commerce while simultaneously creating new vulnerabilities for financial fraud. The increasing sophistication and frequency of fraudulent activities on e-commerce websites, pose significant challenges to the seller and buyer alike. This results in substantial economic losses across the global financial ecosystem. To address this abuse, this paper proposes a hybrid approach combining Knowledge Graphs constructed with embeddings generated on the IEEE-CIS fraud dataset and a novel evidence-based scoring system for online transactions. The embeddings capture both temporal and behavioral patterns in transaction data through time-delta features and other transaction characteristics such as distance, transaction amounts and bank information. This rich representation, combined with our novel Rate of Variable Change (RVC) analysis through the knowledge graph, enables detection of suspicious behavioral patterns in online transactions and assigning an evidence-based risk score. In this paper, we aim to increase recall score for fraudulent transactions, since the cost of a missing fraud transaction is considered much higher than the cost of a false positive.