Feature Selection for Document Ranking using Best First Search and Coordinate Ascent

Feature selection is an important problem in machine learning since it helps reduce the number of features a learner has to examine and reduce errors from irrelevant features. Even though feature selection is well studied in the area of classification, this is not the case for ranking algorithms. In this paper, we propose a feature selection technique for ranking based on the wrapper approach used in classification. Our method uses the best first search strategy incrementally to partition the feature set into subsets. Features in each subset are then combined into a single feature using coordinate ascent in such a way that it maximizes any defined retrieval measure on a training set. Our experiments with many state-of-the-art ranking algorithms, namely RankNet, RankBoost, AdaRank and Coordinate Ascent, have shown that the proposed method can reduce the original set of features to a much more compact set while at least retaining the ranking effectiveness regardless of the ranking method in use.