Somya Rai and MD Sharique argue that a transaction's connections can reveal fraud signals that tabular features miss. Their approach feeds graph embeddings into XGBoost and compares results separately for physical and online transactions at a fixed alert budget. They report relative detection gains, but projected annual fraud capture is not independently verified realized savings.
Somya Rai and MD Sharique describe sampling that preserves user histories, time-respecting edges, separate monthly graphs and training-only feature transforms. Their two-hop aggregation combines mean, maximum and standard deviation so a risky neighbor is not diluted by an otherwise ordinary group. Attribution over linked transactions supplies evidence for analysts and model debugging; it is not a guarantee that every alert is correct.
Somya Rai and MD Sharique separate offline graph computation from a small online scoring path. They report low millisecond latency while explaining the memory cost and stale-embedding performance loss of caching every transaction node. An edge-oriented approach caches users and merchants instead, reducing the footprint. The reported benchmarks and scaling choices describe this system's tests, not universal performance for every fraud pipeline.
Watch on YouTube
