多查询相关的排序支持向量机融合算法

排序学习是目前信息检索与机器学习领域研究的热点问题.现有排序学习算法在学习时把训练样本集中的所有查询及其相关文档等同对待，忽视了查询之间的差异，影响了排序模型的性能.对查询之间的差异进行描述，并在训练过程中考虑这种差异，提出一种基于有监督学习的融合多个与查询相关排序子模型的方法.该方法为每一个查询及其相关文档建立一个子排序模型，并将子排序模型的输出进行向量化表示，将多个查询相关的排序模型转化为体现查询差异的特征数据，实现多排序模型的集成.以排序支持向量机为例，在查询级和样本级建立新的损失函数作为优化目标，并利用此损失函数调节不同查询产生损失之间的权重，提出多查询相关的排序支持向量机融合算法.在文档检索和网页检索中的实验结果表明，使用多查询相关的排序支持向量机融合算法可以取得比传统排序学习模型更好的性能.

A Multiple Query Dependent Ranking SVM Aggregation Algorithm

Supervised ranking approaches have been becoming more and more important in the fields of information retrieval and machine learning. In ranking for document retrieval, queries often vary greatly from one to another. Only the documents retrieved from the same query are to be ranked against each other. However, in most of the existing approaches, losses from different queries are defined as the same. The significant diversities existing among queries are taken into consideration, and a rank aggregation framework for multiple dependent queries is proposed. This framework contains two steps, training of base rankers and query-level aggregation. Training of base ranker sets up a number of query-dependent base rankers based on each query and its relevant documents, and then turns the output of base rankers into feature vectors. Query-level aggregation uses a supervised approach to learn query-dependent weights when these base rankers are combined. As a case study, an SVM based model is employed to aggregate the base rankers, referred to as Q.D.RSVM. It is proved that Q.D.RSVM can set up query-dependent weights for different base rankers. Q.D.RSVM is applied to document retrieval and Web retrieval tasks. Experimental results based on benchmark datasets show that Q.D.RSVM outperforms conventional ranking approaches.