Expected Divergence based Feature Selection for Learning to Rank
Parth Gupta, Paolo Rosso · 2013
Feature selection methods are essential for learning to rank (LTR) approaches as the number of features are directly proportional to computational cost and sometimes, might lead to the over-fitting of the ranking model. We propose an expected divergence based approach to select a subset of highly discriminating features over relevance categories. The proposed method is evaluated in terms of performance of standard LTR algorithms when trained with reduced features over a set of standard LTR datasets. The proposed method leads to not significantly worse, and in some cases, significantly better performance compared to the baselines with as few features as less than 10%. The proposed method is scalable and can easily be parallelised.