Papers and Presentations 7 KEYNOTE — The DDI Approach to Features : Don ’ t Do It

Ranking is a central problem in information retrieval. Much work has been done in the recent years to automate the development of ranking models by means of supervised machine learning. Feature selection aims to provide sparse models which are computationally efficient to evaluate, and have good ranking performance. We propose integrating the feature selection as part of the training process for the ranking algorithm, by means of a wrapper method which performs greedy forward selection, using leave-query-out crossvalidation estimate of performance as the selection criterion. We introduce a linear time training algorithm we call greedy RankRLS, which combines the aforementioned procedure, together with regularized risk minimization based on pairwise least-squares loss. The training complexity of the method is O(kmn), where k is the number of features to be selected, m is the number of training examples, and n is the overall number of features. Experiments on the LETOR benchmark data set demonstrate that the approach works in practice.