Judge a Transaction by the Company It Keeps — Somya Rai & MD Sharique, EXL

A transaction can look harmless in isolation and suspicious once its links to recent fraud become visible. Somya Rai and MD Sharique explain an EXL fraud detection system that adds graph embeddings to an existing XGBoost classifier. The objective is precision within a fixed analyst alert budget, rather than a better headline score alone. Rai introduces the neighborhood signal and the constraints of two transaction portfolios; Sharique then walks through sampling, graph construction, explanation and serving. They report relative detection improvements of 27% to 33% against their tabular baseline at matched alert rates. The financial figure discussed is projected annualized incremental fraud capture, rather than realized savings. The engineering decisions start before training. Selected users retain their transaction histories so sampling does not destroy the neighborhoods the model needs. Separate monthly graphs, earlier timestamps and training only preprocessing constrain leakage. GraphSAGE combines mean, maximum and standard deviation, preserving an unusual neighbor that averaging alone could obscure. Attribution connects a suspicious score to specific transaction edges for analyst review. For serving, cached representations and a small online forward pass keep graph sized work off the request path. Moving from transaction node embeddings to user and merchant embeddings reduces the stated cache footprint from roughly 60 GB to 2 to 3 GB. The talk closes with rolling windows, temporal decay and reported millisecond latency, showing the tradeoffs behind the system rather than presenting a universal fraud benchmark. Speaker info: Somya Rai: - https://www.linkedin.com/in/somyarai/ MD Sharique: - https://www.linkedin.com/in/mdsharique0107/ Timestamps: 00:00 - Somya Rai: fraud signals in transaction neighborhoods 2:09 - Why more tabular features hit a ceiling 3:12 - Learning from a transaction neighborhood 5:02 - Four engineering decisions 6:20 - MD Sharique: sampling without breaking connectivity 8:29 - Monthly graphs and leakage controls 9:32 - Entity history and temporal features 11:06 - GraphSAGE with mean, max and standard deviation 13:37 - Comparing detection at fixed alert rates 14:32 - Edge attribution for analyst review 16:40 - Cached representations and online scoring 18:26 - Moving from node to edge prediction