Problem
Fraud attacks change constantly, and attackers coordinate on social media. A model trained on last month's patterns may not catch a genuinely new one, and a new attack usually comes with very little labeled data, sometimes a handful of confirmed cases. A deny rule is the fastest fix. But finding one that stops the attack without blocking good customers takes a human half a day to a full day. Hand-picked rules are often too broad (good customers blocked) or too narrow (routed around within days).
Example from the article: a fictionalized account-takeover pattern: a burst of purchases on newly added shipping addresses, right after a password reset, from a device the account has never used.
Solution in brief
Treat rule writing as a search problem. For an analyst-chosen merchant or time window, a pipeline gathers data and qualitative notes, searches for discriminating signals, and drafts a candidate rule with precision, recall and customer-impact estimates. A human analyst reviews and approves it.
Questions this case answers
- How to find a precise, explainable decision rule for a new adversarial pattern when only a handful of labeled cases exist and a trained model would overfit
- How to speed up expert hand-crafting of decision rules by automating signal search while keeping a human approval step
- How to write a fraud rule for a brand-new attack pattern in minutes instead of a day?
- How to make a hand-written deny rule neither too broad nor too narrow?
Tools and methods
- Signal screening — ranks each candidate signal on its own
- Pattern mining — extracts structure from free-text fields
- Supervised rule induction — grows and prunes rules to find distinct fraud sub-populations
- AI drafting layer — combines the best structures with attack context, sweeps thresholds, filters overfit signals and writes the rule with its rationale
- Analyst review — final human check
Google or a one-time email code. No password.