An Ensemble Approach to Learning to Rank

In recent years, ´learning to rank´ is a focused approach for information retrieval, which can learn the ranking order given by experts and construct a uniform model to rank for new query. But in practice user queries vary in large diversity, it makes a single learned ranker not representative. Therefore, we propose an ensemble approach to ´learning to rank,´ in which a lower generalization error can be gotten by generating a set of rankers and leveraging these rankers for the final prediction. Moreover, two strategies of creating multiple base rankers are proposed to make the ensemble more effective for information retrieval. The experiment results on two real world datasets indicate that the proposed approach can outperform the original ´learning to rank´ methods significantly.