Learning to Rank for Cross-Device Identification

We summarize the 1st place solution of the ICDM 2015 data mining contest and explain how the goal of the contest fits naturally into the learning-to-rank paradigm. We show that learning-to-rank models provide an improvement over the more conventional binary classification models for this task. Finally, we explain our gradient weighting and cookie over-selection strategies which provided some additional lift to our final score.