RankNet

RankNet is a neural network-based learning-to-rank algorithm developed by researchers at Microsoft Research in the mid-2000s. It is designed to learn a scoring function that can order documents with respect to a given query by modeling pairwise preferences between documents.

The core idea is to assign a real-valued score s for each document using a shared neural network f(x; θ) over its feature vector x. For a pair of documents i and j, the model defines the probability that i should be ranked above j as P(i > j) = sigmoid(s_i − s_j). The ground truth provides a binary label y_{ij} indicating whether i is preferred to j. The training objective is the cross-entropy loss over pairs: L_{ij} = −[ y_{ij} log P(i > j) + (1 − y_{ij}) log(1 − P(i > j)) ]. The network is trained by backpropagation to minimize the sum of these pairwise losses across selected document pairs, typically constructed from query-labeled data.

RankNet popularized end-to-end neural approaches to ranking by using a differentiable surrogate loss and pairwise comparisons, rather than relying solely on traditional ranking signals. It laid the groundwork for subsequent learning-to-rank methods that extend pairwise ideas to listwise or more scalable frameworks, such as LambdaRank and related neural ranking models.

Applications of RankNet include information retrieval and web search ranking, where the goal is to order results to maximize user satisfaction. While the original formulation focuses on pairwise preferences, the concept has influenced broader neural ranking research, contributing to the development of more scalable and metric-aligned approaches in practice.