ERR.Rank: An algorithm based on learning to rank for direct optimization of Expected Reciprocal Rank

摘要

Learning to rank (LTR) is a machine learning-based ranking technique that constructs a ranking model to sort objects in response to a query, and is used in many applications especially in information retrieval. LTR ranking models are generally evaluated using information retrieval measures. Listwise approaches are among the most important learning to rank algorithms. A subset of listwise approaches try to optimize the evaluation measures. These evaluation measures are dependent only on the document positions in the ranking and are discontinuous and non-convex with respect to the scores of the ranking function. The majority of evaluation measures used by current listwise techniques ignore the relationship between a document at a position and the documents at higher positions. To overcome this problem, we propose a new listwise algorithm, which aims to directly optimize the Expected Reciprocal Rank (ERR) measure. ERR considers the importance of a document at a position to be dependent on the documents ranked higher than this document. Our algorithm uses a probabilistic framework to optimize the expected value of ERR. We use a boosting approach using a gradient descent in order to find the optimal ranking function. The presented algorithm is compared with state of the art algorithms. The results obtained on the ’LETOR 3.0’ standard dataset indicate that the proposed method outperforms the baselines.