RankNet

class lightning_ir.loss.pairwise.RankNet(temperature: float = 1)

Bases: PairwiseLossFunction

RankNet loss function for pairwise ranking tasks.

RankNet optimizes pairwise ranking by modeling the probability that a positive document should be ranked higher than a negative document using a logistic function. It computes the margin between the scores of positive and negative pairs and applies a binary cross-entropy loss to maximize the likelihood of correct pairwise orderings. This approach allows the model to learn from relative comparisons rather than absolute score values.

Originally proposed in: Learning to Rank using Gradient Descent

__init__(temperature: float = 1) → None

Initialize the RankNet loss function. :param temperature: Temperature parameter for scaling the scores. :type temperature: float

Methods

__init__([temperature])	Initialize the RankNet loss function.
compute_loss(output, batch)	Compute the RankNet loss.

compute_loss(output: LightningIROutput, batch: TrainBatch) → torch.Tensor

Compute the RankNet loss.

Parameters:

• output (LightningIROutput) – The output from the model containing scores.

• batch (TrainBatch) – The training batch containing targets.

Returns:

The computed loss.

Return type:

torch.Tensor