About Anti-Fraud and Cats: How to Fight Fraud, Not Just Detect It Better

Aleksei Nikiforov23 May 2026

Problem

A platform lets owners and users add objects (organizations on Yandex Maps), and fraudsters add fake ones to get customers, promote sites or scam people. Between obvious fraud and obvious honest objects lies a large grey zone where intent cannot be seen. Hundreds of thousands of attempts are banned every year, and in Moscow and St. Petersburg fraud sometimes outweighs honest submissions. The cost is cluttered maps, lost trust and real losses for users.

Example from the article: fake visiting-vet clinics on Yandex Maps. A pet owner in an urgent situation opens a card, calls, summons a vet home without asking the price, and in one screenshot is charged 8 thousand rubles, not the maximum.

Solution in brief

The goal of anti-fraud can be set three ways: classify as accurately as possible, make fraud unprofitable, or guarantee zero fraud at any price. Each goal gives different tools and costs. Yandex rolled out most of the controls across all maps and ran a zero-tolerance track only for vet clinics.

Questions this case answers
  • How to choose the objective of a detection system against adaptive bad actors when a large grey zone exists: classify more accurately, make attacks unprofitable, or enforce zero tolerance for one category
  • How to build features for a newly created object with almost no history and a golden set from past attacks and moderation verdicts
  • How to choose the goal of an anti-fraud system when fraudsters adapt and there is a large grey zone
  • How to make fraud unprofitable instead of only classifying it more accurately

Tools and methods

  • Baseline — block only obvious spam, leave the rest to moderation
  • Classification with a golden set — past attacks, moderation verdicts, retrospective labels
  • Graph features — bipartite user-object graph; object graph with neighbor aggregations
  • Fraud economics — control tuned to visibility and demand, faster bans of phones and websites
  • Zero tolerance — whole grey zone treated as fraud, mandatory manual verification
Continue reading — it's free

Google or a one-time email code. No password.