Learning to Rank using Gradient Descent

  • Chris J.C. Burges
  • Tal Shaked
  • Erin Renshaw
  • Ari Lazier
  • Matt Deeds
  • Nicole Hamilton
  • Greg Hullender

MSR-TR-2005-06 |

We investigate using gradient descent methods for learning ranking functions; we propose a simple probabilistic cost function, and we introduce RankNet, an implementation of these ideas using a neural network to model the underlying ranking function. We present test results on toy data and on data from a commercial internet search engine.