RankNet

RankNet is a seminal Pairwise Ranking algorithm within the broader field of Learning to Rank (LTR). It was primarily developed by Chris Burges and his colleagues at Microsoft Research. The algorithm utilizes a Neural Network to map a feature vector of an item to a real-valued score, which is then used to determine its position in a ranked list.

The core innovation of RankNet is its probabilistic cost function. Instead of optimizing for the absolute position of an item, the model learns the relative order of pairs of items. For any given pair, the model predicts the probability that one item should be ranked higher than the other using a Sigmoid Function applied to the difference of their scores. The training process involves minimizing the Cross-Entropy loss between the predicted probabilities and the ground truth labels. This optimization is typically performed using Stochastic Gradient Descent. Detailed technical specifications are available in the original paper, Learning to Rank using Gradient Descent.

While RankNet was a significant advancement, it later evolved into more sophisticated algorithms like LambdaRank and LambdaMART. These successors were designed to directly optimize non-differentiable Information Retrieval metrics such as NDCG and Mean Average Precision. A historical overview of these developments can be found on the Wikipedia page for Learning to Rank.

• LambdaRank
• LambdaMART
• Information Retrieval
• Gradient Descent