Search engines must order results, not just classify them as relevant or not, and the quality that users notice is whether the best results sit at the top. In 2005 Chris Burges and colleagues at Microsoft Research presented RankNet at the International Conference on Machine Learning, introducing a practical way to train a neural network directly on the ranking problem.
RankNet’s trick is to learn from pairs of items. Instead of predicting an absolute score, the model is told that document A should rank above document B for a given query, and it learns a scoring function whose differences match those preferences. The authors used a probabilistic cost function based on the difference between two items’ scores, turning the ordering signal into a smooth objective that ordinary gradient descent can optimize. The model itself was a neural network, but the framework works with any differentiable scorer.
RankNet was deployed in commercial web search and became the seed of a line of algorithms. Its successors, LambdaRank and the tree-based LambdaMART, went on to win the 2010 Yahoo Learning to Rank Challenge and to power production ranking at major search engines. The 2005 paper is the foundational reference for the shift from hand-tuned ranking formulas to ranking functions learned from data.
For a business reader, this is the moment search ranking became a machine learning problem: the order of the links you see is increasingly the output of a model trained on millions of human preference judgments.