lenskit.reranking.fair.FAIRReranker

class lenskit.reranking.fair.FAIRReranker(config=None, **kwargs)

Bases: lenskit.pipeline.Component[lenskit.data.ItemList], lenskit.training.Trainable

FA*IR top-N re-ranking algorithm.

Re-ranks a sorted candidate list to satisfy binomial lower-bound constraints for a binary protected/unprotected attribute at every prefix (1..N). It expects a sorted list of items, the target proportion of protected items p and significance level alpha.

For details, see Zehlike et al. (2017).

Stability:

Caller (see Stability Levels).

Parameters:

• config (object | None)

• kwargs (Any)

config: FAIRRerankerConfig

FA*IR reranker configuration.

alpha_c: float

m_list: list[int]

is_trained()

Query if this component has already been trained.

train(data, options=TrainingOptions())

Precompute adjusted alpha and m-table for the configured (n,p,alpha)

Training computes the (multiple-tests–corrected) prefix requirements m[i] so that every prefix of size i has at least m[i] protected items with significance α (after correction).

Parameters:

• attributes (protected) – The item entity table must have a scalar attribute protected that is True if the item is in the protected group and False otherwise.

• data (lenskit.data.Dataset)

• options (lenskit.training.TrainingOptions)

__call__(items, n=None)

Run the pipeline’s operation and produce a result. This is the key method for components to implement.

Parameters:

• items (lenskit.data.ItemList)

• n (int | None)

Return type:

lenskit.data.ItemList