Listwise approach to learning to rank: theory and algorithm
Proceedings of the 25th international conference on Machine learning
Learning to Rank for Information Retrieval
Foundations and Trends in Information Retrieval
Learning to rank with cross entropy
Proceedings of the 20th ACM international conference on Information and knowledge management
| Hi-index | 0.00 |
An essential issue in document retrieval is ranking, and the documents are ranked by their expected relevance to a given query. Multiple labels are used to represent different level of relevance for documents to a given query, and the corresponding label values are used to quantify the relevance of the documents. According to the training set for a given query, the documents can be divided into several groups. Specifically, the documents with the same label are assigned to the same group. If the documents in the group with higher relevance label can always be ranked higher over the ones in groups with lower relevance label by a ranking model, it is reasonable to expect perfect ranking performance. Inspired by this idea, we propose a novel framework for learning to rank, which depends on two new samples. The first one is one-group constituted by one document with higher level label and a group of documents with lower level label; the second one is group-group constituted by a group of documents with higher level label and a group of documents with lower level label. A novel loss function is proposed based on the likelihood loss similar to ListMLE. We demonstrate the advantages of our approaches on the Letor 3.0 data set. Experimental results show that our approaches are effective in improving the ranking performance.