Group-enhanced ranking

An essential issue in document retrieval is ranking, which is used to rank documents by their relevancies to a given query. This paper presents a novel machine learning framework for ranking based on document groups. Multiple level labels represent the relevance of documents. The values of labels are used to quantify the relevance of the documents. According to a given query in the training set, the documents are divided into several groups based upon their relevance labels. The group with higher relevance labels is always ranked upon the ones with lower relevance labels. Further a preference strategy is introduced in the loss functions, which are sensitive to the group with higher relevance labels to enhance the group ranking method. Experimental results illustrate that the proposed approach is very effective, with a 14 percent improvement on TD2003 data set evaluated by MAP.