Multi-layer Classification: ICDM 2015 Drawbridge Cross-Device Connections Competition

摘要

This paper outlines a numerical recipe used to solve the ICDM 2015 competition to associate devices and cookies. The presented model utilizes binomial classification at three stages. The first stage imposed a shared IP address constraint to produce an initial set of candidate matches. The second stage utilized gradient boosted machines with 90 calculated features. This stage was the focal point for solving the problem. The final classification reassessed the probabilities of the prior stage and was customized to fit specific competition dynamics. Our team finished in 10th place in the public and private leaderboard with an F0.5 score of 0.855150.