By 2010 the family of learning-to-rank algorithms that grew out of RankNet had spread across the details of several papers and internal reports. Chris Burges wrote this Microsoft Research technical report (MSR-TR-2010-82) to pull them together into one self-contained explanation of RankNet, LambdaRank, and LambdaMART.
The thread connecting the three is a problem with ranking metrics. The quality measures that matter, such as NDCG, depend on the sorted order of results and are not smooth, so they cannot be optimized directly by gradient descent. RankNet sidesteps this by optimizing a pairwise cost. LambdaRank’s insight is that you do not need the cost function itself, only its gradients, called lambdas, which can be defined to directly reflect how much swapping two results would change the target metric. LambdaMART then plugs those lambdas into gradient-boosted decision trees instead of a neural network, combining the metric-aware gradients with a powerful ensemble model.
LambdaMART became one of the strongest off-the-shelf ranking methods of its era; a LambdaMART ensemble won the 2010 Yahoo Learning to Rank Challenge. The report is the standard reference that explains, in one place, how production search ranking actually optimizes for the metrics users care about.
For a business reader, this document is the bridge from research idea to deployed system: it is the practical recipe that ranking teams reached for when they needed search results ordered to maximize a real quality metric.