Active learning for ranking with sample density

摘要

While ranking is widely used in many online domains such as search engines and recommendation systems, it is non-trivial to label enough data examples to build a high performance machine-learned ranking model. To relieve this problem, active learning has been proposed, which selectively labels the most informative examples. However, data density, which has been proven helpful for data sampling in general, is ignored by most of the existing active learning for ranking studies. In this paper, we propose a novel active learning for ranking framework, generalization error minimization (GEM), which incorporates data density in minimizing generalization error. Concentrating on active learning for search ranking, we employ classical kernel density estimation to infer data density. Considering the unique query-document structure in ranking data, we estimate sample density at both query level and document level. Under the GEM framework, we propose new active learning algorithms at both query level and document level. Experimental results on the LETOR 4.0 data set and a real-world Web search ranking data set from a commercial search engine have demonstrated the effectiveness of the proposed active learning algorithms.