LambdaRank
- Key observations of LambdaRank:
- To train a model, we do not need the costs themselves, only the gradients (of the costs. wrt model scores).
- The gradient should be bigger for pairs of documents that produces a bigger impact in NDCG by swapping positions.
LambdaRank [Burges et al., 2006] Multiply actual gradients with the change in NDCG by swapping the rank positions of the two documents:
$$
\lambda_{\text {LambdaRank }}=\lambda_{\text {RankNet }} \cdot|\Delta \mathrm{NDCG}|
$$
-
This approach also works with other metrics, e.g. $\mid \Delta$ Precision|
Empirically LambdaRank was shown to directly optimize IR metrics. -
Recently, it was theoretically proven that LambdaRank optimizes a lower bound on certain IR metrics [Wang et al., 2018 ]