| 摘要: |
| 针对当前基于支持向量机的排序学习方法训练时间长以及不考虑查询之间差异、模型单一的问题,提出一种查询依赖的有序多超平面排序学习模型.根据不同查询,利用其对应训练数据所属等级之间的序关系构建多个超平面.此外,提出了一种加权表决方法对多个超平面的排序列表进行聚合,根据各超平面的排序精度赋予其不同权重,计算最终排序结果.在标准数据集LETOR OHSUMED 上对所提出的模型性能进行了综合评测,并与相关排序模型进行了对比分析.实验结果显示,所提出的模型排序性能有较大提升.同时,训练时间明显缩短. |
| 关键词: 排序学习 查询依赖 多超平面 序关系 聚合 加权表决 |
| DOI：10.3724/SP.J.1001.2011.03908 |
| 分类号: |
| 基金项目:国家高技术研究发展计划(863)(2008AA01Z131); 陕西省自然科学基础研究计划(SJ08-ZT14) |
|
| Query Dependent Learning to Model Based on Ordered Multiple Hyperplanes |
|
SUN He-Li1, HUANG Jian-Bin2, FENG Bo-Qin1, ZHAO Zhi-Qin1, LIU Jun1, ZHENG Qing-Hua1
|
|
1.Department of Computer Science and Technology, Xi’an Jiaotong University, Xi’an 710049, China;2.School of Software, Xidian University, Xi’an 710071, China
|
| Abstract: |
| This paper proposes a ranking model that trains different hyperplanes for different queries and optimizes hyperplanes with the order relations. It aims at solving the problem of most existing rank methods that do not consider the significant differences between queries and only resort to a single function that is time consuming. Next, a weighted voting method is proposed to aggregate the ranking lists of the hyperplanes as the final rank. The weights reflect the degree of precision. Effectiveness is tested by the benchmark data set LETOR OHSUMED and is compare with other ranking models. The proposed method shows improved ranking performance with a significant reduction of training time. |
| Key words: learning to rank query dependent multiple hyperplane order relation aggregation weighted voting |
