Judge a Transaction by the Company It Keeps — Somya Rai & MD Sharique, EXL
Fraud is a property of transaction environment, not the transaction itself
“The signal was not in the string itself, but in the connections. Therefore, fraud is not a property of the transaction, but of the environment in which it occurs.”
EXL's AI team demonstrates that adding transaction context (neighboring fraudulent activity by same user/merchant) can shift a fraud score from 0.04 to 0.9 on an identical feature vector, proving relational signals outperform tabular feature engineering. They show that doubling feature counts (40→87, then 200) yields only ~1 percentage point TPR gain, hitting a fundamental ceiling of single-card tabular data. The engineering insight is that fraud detection requires graph-aware or context-aware models, not just better XGBoost on richer feature sets.