Hybrid learning framework for web information retrieval
Guang Feng, Kin‐Man Lam, Xudong Zhang, Desheng Wang · 2008
Machine learning techniques have been considered a very promising solution to Web information retrieval, which is based on the ranking of the relevance of samples to a query input. However, the connotation of labeling in ranking is quite different from that in classification. Specifically, the labeling of samples for ranking is usually incomplete, i.e. only a part of samples are labeled. In order to remedy this methodological gap, in this paper we propose a hybrid learning framework, called fuzzy-label learning, which consists of two layers. First, we utilize a label-propagation algorithm to estimate those labels of unlabeled samples by their neighborhoods. Second, we adopt RankBoost on the samples with fuzzy labels. Experiments with five-fold cross-validation using the Letor benchmark datasets show that the proposed hybrid learning framework can definitively improve the search performance achieved by the RankBoost algorithm for Web information retrieval.