Robust Ranking Models Using Noisy Feedback
Christian Pölitz, Ralf Schenkel · MPG.PuRe (Max Planck Society) · 2012
Direct feedback of users of search engines by click information is naturally noisy. Ranking models that integrate such feedback in their training process must cope with this noise. In worst case such noise can lead to large variance among the results for different queries in the resulting rankings. We propose to integrate model averaging like bagging and random forest methods to reduce the variance in the ranking models. We perform an experimental study on different noise levels using a state of the art ranking model.