Abstract

The problem of ranking the documents according to their relevance to a given query is a hot topic in information retrieval. Most learning-to-rank methods are supervised and use human editor judgements for learning. In this paper, we introduce novel pairwise method called YetiRank that modifies Friedman’s gradient boosting method in part of gradient computation for optimization and takes uncertainty in human judgements into account. Proposed enhancements allowed YetiRank to outperform many state-of-the-art learning to rank methods in offline experiments as well as take the first place in the second track of the Yahoo! learning-to-rank contest. Even more remarkably, the first result in the learning to rank competition that consisted of a transfer learning task was achieved without ever relying on the bigger data from the “transfer-from” domain.