Detect Suspicious Value Transfers in Poker is a machine learning competition challenging participants to build models that identify coordinated gameplay patterns within 2,000,000 synthetic poker hands. The goal is to rank evaluation pairs, identify the likely behavior, and submit specific hands that best support each alert.
The Challenge Suspicious play can often be explained by innocent factors such as tilt, inexperience, repeated schedules, unusual strategy, or simple luck. Coordinated players may also act only occasionally, mixing manipulated hands with normal play. Participants must build a system that separates these cases. Submissions are required to detect suspicious pairs and point to concrete hands an investigator could review.
Target Behaviors
The competition focuses on detecting the following behaviors:
* Directed transfer: One player deliberately loses value to another.
* Soft play: Partners avoid normal aggression against each other.
* Coordinated isolation: Partners pressure other players while limiting conflict between themselves.
* Other coordination: A fourth, undisclosed mechanism that never appears in public positive labels. This should be submitted as other_coordination when detection fits none of the three disclosed patterns.
Coordination is episodic; a coordinated pair mixes manipulated hands with ordinary play, and a relationship is not necessarily active over the whole timeline. The dataset also includes difficult non-target patterns such as tilt, weak play, similar strategies, repeated opponent selection, streaks, and strategy changes.
Data Overview The release contains 2,000,000 six-player No-Limit Texas Hold'em hands involving 12,000 synthetic players. Players are organized into 400 persistent pools of 30, playing at one table for about 5,000 hands. The first 60% of hands form the development period, and the final 40% form the evaluation period.
Prizes The total prize pool is $5,000: * 1st place: $3,000 * 2nd place: $1,250 * 3rd place: $750
To remain eligible for prizes, winning teams must publish a Kaggle Solution Writeup (max 1,500 words), a public notebook or repository containing the code needed to reproduce the submission, and five short case reviews from the submitted evidence within seven days of the Private Leaderboard release.