Abstract

This paper describes the ideas and methodologies that we used in the Yahoo learning-to-rank challenge¹. Our technique is essentially pointwise with a listwise touch at the last combination step. The main ingredients of our approach are 1) preprocessing (querywise normalization) 2) multi-class AdaBoost.MH 3) regression calibration, and 4) an exponentially weighted forecaster for model combination. In post-challenge analysis we found that preprocessing and training AdaBoost with a wide variety of hyperparameters improved individual models significantly, the final listwise ensemble step was crucial, whereas calibration helped only in creating diversity.