Learning to Rank using an Ensemble of Lambda-Gradient Models

We describe the system that won Track 1 of the Yahoo! Learning to Rank Challenge.

The Yahoo! Learning to Rank Challenge, Track 1, was a public competition on a Machine Learning for Information Retrieval task: given a set of queries, and given a set of retrieved documents for each query, train a system to maximize the Expected Reciprocal Rank (Chapelle et al., 2009) on a blind test set, where the training data takes the form of a feature vector x ∈ Rd with label y ∈ Y, Y ≡ {0,1,2,3,4} (a more positive number denoting higher relevance) for each query/document pair (the original, textual data was not made available). The Challenge setup, background information, and results have been extensively covered elsewhere and we refer to Chapelle and Chang (2011) for details. In this paper we summarize the work that resulted in the winning system.1 We limit the work described in this paper to the work done specifically for the Challenge; the work was done over a four week period prior to the end of the Challenge.