RecallGain metric

[blondered]

Feature Description

Given test interactions, we can exclude all Positives that would have been recommended by the reference model (e.g. popular model). After that we can calculate Recall as usual. Resulting metrics is interpreted as a gain to overall Recall which is achieved by the new algorithm in comparison with the reference algorithm.

Why this feature?

Useful for both candidate-generators selection and simple validation protocol in case of strong popularity bias in data (e.g. calculating RecallGain from a Popular model).

Additional context

recall_gain = RecallGain(k=10, k_ref=20)

# one metric
value = recall_gain.calc(reco=recos, ref_recos=ref_recos)
per_user = recall_gain.calc_per_user(reco=recos, ref_recos=ref_recos)

# calc_metrics
calc_metrics(
    metrics = {"precision": precision, "intersecton": intersecton, "recall_gain": recall_gain},
    reco=recos,
    interactions=df_test,
    ref_recos: Union[pd.Datafame, Dict[Hashable, pd.DataFrame]]=ref_recos
)
# here we can pass multiple models with:
ref_recos = {"one": ref_recos_one, "two": ref_recos_two}
# result dict will have `intersecton_one`, `intersecton_two`, `recall_gain_one`, `recall_gain_two` as keys

# cross_validate
cv_results = cross_validate(
    dataset=dataset,
    splitter=splitter,
    models=models,
    metrics=metrics,
    k=10,
    filter_viewed=True,
    ref_models=["one", "two"], # just selecting keys from `models` argument
    validate_ref_models=False  # optionally exclude ref_models from other metrics
)