实验室关于排序学习研究成果被期刊Expert Systems with Application录用
新闻来源:IR实验室       发布时间:2019/4/30 21:55:12

  近日, 收到期刊《Expert Systems with Application》编辑部邮件,实验室林原老师等的研究工作“FGFIREM: A Feature Generation Framework based on Information Retrieval Evaluation Measures”已被录用。

  摘要:排序学习是信息检索领域的重要研究内容,近年来,很多排序学习方法有效地改善了检索性能。本文基于AdaRank排序学习方法直接优化信息检索评价指标,通过直接优化信息检索评价指标ERR,MRR和Q-measure提出三种排序学习新方法。进而,提出一种新的排序特征生成框架FGFIREM,通过拓展排序学习特征空间,进一步优化排序性能。相关方法在排序学习标准数据集LETOR3.0和MSLR-WEB10K上均有效提升了排序性能。

  Abstract:Learning to rank has become one of the most popular research areas in recent years. A series of learning to rank algorithms have been proposed to improve the ranking performance. In this work, we propose three learning to rank algorithms by directly optimizing evaluation measures based on the AdaRank algorithms.  We name the three algorithms as AdaRank-ERR , AdaRank-MRR and AdaRank-Q, which optimize three evaluation measures, Expected Reciprocal Rank (ERR), Mean Reciprocal Rank (MRR), and Q-measure (Q), based on AdaRank, respectively. Furthermore, we propose a novel feature generation framework FGFIREM to enhance the ranking performance. The framework generates effective document ranking features based on the ranking scores assigned by the proposed algorithms, and enriches the original feature space of learning to rank using the generated features for improving the ranking performance.  We evaluate the proposed framework on three datasets from LETOR3.0 and the web dataset MSLR-WEB10K. The experimental results demonstrate that our framework can effectively improve the ranking performance.