Epits for Relational Feature Engineering: Evidence from Financial Transaction Networks

Many prediction problems depend on relationships among entities and how those relationships change over time. The challenge is deciding which parts of that structure and history to preserve as model features. We investigate epits as a framework for relational feature engineering. An epit has three positions and three pairwise relationships; their roles suggest interpretable measurements of direction, memory, change, and surrounding activity. We test this general approach in financial transaction networks, using synthetic IBM bank payments with transaction-level laundering labels and observed Ethereum transfers with retrospective phishing marks on receivers. On IBM, sixteen epit-guided features raise later-period average precision (AP) from 0.352 to 0.606 over a baseline already containing the Graph Feature Preprocessor (GFP). AP measures how well labeled cases are ranked near the top, not classification accuracy. Direct sender–receiver histories explain much of that gain. Yet the sixteen features still add 3.05 AP points when both direct-pair histories and a conventional 30-day account-history panel are already provided. A separate test after direct-pair history finds the largest further contribution in measures of prior activity and how much evidence supports a relationship, rather than intricate pattern changes. On Ethereum, using all earlier sender contacts improves on recency features restricted to selected recent counterparties. Together, the tests show how epit-guided exploration can uncover useful, traceable relational measurements, including information missing from the tested GFP baseline on IBM. Epits offer a way to discover these features, not the only way to compute them. The results are exploratory and do not establish superiority over temporal graph neural networks.