Research On Learning To Rank Models Of Directly Optimizing Evaluation Measures

Ranking is an important issue related to user experience in information retrieval, how to rank the retrieved result correctly has gained more and more attention. As a new research focus, learning to rank plays a very important role in page ranking.One of the central issues in learning to rank is to develop algorithms that construct ranking models by directly optimizing evaluation measures used in information retrieval such as Mean Average Precision (MAP) and Normalized Discounted Cumulative Gain (NDCG). Several such algorithms have been proposed and their effectiveness has been verified.In learning to rank, we need a lot of labeled data as our training set, it is costly to label documents for different research areas. Can we construct a general model that can be used as a cross domain model to make prediction in domain A while trained using data from domain B. Some efforts are made and some results are achieved. In this paper, we propose a novel method which can be used as a cross-domain adaptive model based on importance weighting. We can get different result for different test data.The evaluation measures used in information retrieval are all discontinuous, many researchers wish to optimize a smooth and differentiable upper bound of the evaluation measure. In this paper, we just optimize the non-smooth objective directly, inspired by AdaBoost and AdaRank, a transductive ranking model named IW-AdaRank is proposed. Experiments on OHSUMED show that our method performs better than some other state-of-the-art methods.Evolutionary algorithms are randomized search method inspired by biological evolution. Their main characteristic is that the objective functions are not limited to smooth functions. They are highly parallel, with the characteristics of global optimization. They are suitable for optimizing the non-smooth evaluation measures in IR. We propose two evolutionary algorithm-based models named GARank and PSORank, experiments results show that both of them perform very well.