Learning to rank using gradient descent

We investigate using gradient descent meth- ods for learning ranking functions; we pro- pose 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.