[FEATURE] Add fairness and popularity bias evaluation metrics

[bajajahsaas]

Description

The evaluation module currently provides 21 metrics across rating accuracy (RMSE, MAE, R-squared, etc.), ranking quality (Precision@k, Recall@k, NDCG@k, MAP, etc.), and beyond-accuracy dimensions (diversity, novelty, serendipity, coverage). However, it has no metrics for fairness or popularity bias -- two increasingly important evaluation dimensions in recommendation systems research.

Fairness in recommendations is a growing concern with regulatory implications (e.g., the EU AI Act requires fairness assessments for AI systems). Popularity bias is one of the most well-studied biases in recommender systems, where algorithms disproportionately recommend already-popular items at the expense of niche/long-tail content.

I'd like to contribute 10 new metrics across these two categories:

Popularity Bias Metrics (5):

Metric	Description	Reference
Average Recommendation Popularity (ARP)	Average popularity of recommended items	Abdollahpouri et al., FLAIRS 2019
Average Percentage of Long Tail Items (APLT)	Fraction of long-tail items in recommendations	Abdollahpouri et al., FLAIRS 2019
Average Coverage of Long Tail Items (ACLT)	Coverage of long-tail catalog	Abdollahpouri et al., FLAIRS 2019
Popularity Lift	Popularity amplification ratio (reco vs train)	Abdollahpouri, PhD Thesis 2020
Gini Index	Inequality of item recommendation frequency	Standard inequality measure

Fairness Metrics (5):

Metric	Description	Reference
Group Metric Disparity	Meta-metric: any existing metric's gap across user groups	Li et al., WWW 2021
Demographic Parity	Equal recommendation rates across groups	Burke et al., FAccT 2018
Equal Opportunity Difference	Recall@k gap across groups	Adapted from Hardt et al., NeurIPS 2016
Calibration Error	KL divergence between user preferences and reco distribution	Steck, RecSys 2018
Exposure Fairness	Gini of exposure across item providers	Singh and Joachims, KDD 2018

Expected behavior with the suggested feature

• A new file recommenders/evaluation/python_evaluation_fairness.py with Python/pandas implementations of all 10 metrics, following the existing coding conventions (decorators, docstrings with citations, constants from recommenders.utils.constants)
• A new SparkFairnessEvaluation class in spark_evaluation.py with PySpark versions of the popularity bias metrics (following the SparkDiversityEvaluation pattern)
• Unit tests with boundary condition checks (e.g., Gini=0 for uniform distribution, calibration error=0 when preference distribution matches recommendation distribution)
• An example notebook examples/03_evaluate/fairness_and_bias_evaluation.ipynb demonstrating the metrics, comparing a biased vs fair recommender
• Zero new dependencies -- uses only numpy, pandas, and sklearn (already required)

Proposed implementation details:

File	Action
recommenders/utils/constants.py	Modify: add DEFAULT_GROUP_COL, DEFAULT_PROVIDER_COL, DEFAULT_LONG_TAIL_THRESHOLD
recommenders/evaluation/python_evaluation_fairness.py	Create: 10 metric functions + decorators + helpers
recommenders/evaluation/spark_evaluation.py	Modify: add SparkFairnessEvaluation class
tests/unit/recommenders/evaluation/conftest.py	Modify: add fairness_data fixture
tests/unit/recommenders/evaluation/test_python_evaluation_fairness.py	Create: 22 unit tests with boundary checks
tests/unit/recommenders/evaluation/test_spark_evaluation_fairness.py	Create: Spark parity tests
examples/03_evaluate/fairness_and_bias_evaluation.ipynb	Create: example notebook

Willingness to contribute

• Yes, I can contribute for this issue independently.
• Yes, I can contribute for this issue with guidance from Recommenders community.
• No, I cannot contribute at this time.

Other Comments

References:

1. H. Abdollahpouri, R. Burke, B. Mobasher. "Managing Popularity Bias in Recommender Systems with Personalized Re-ranking." FLAIRS 2019
2. H. Steck. "Calibrated Recommendations." RecSys 2018
3. A. Singh, T. Joachims. "Fairness of Exposure in Rankings." KDD 2018
4. R. Burke, N. Sonboli, A. Ordonez-Gauger. "Balanced Neighborhoods for Multi-Sided Fairness in Recommendation." FAccT 2018
5. Y. Li et al. "User-oriented Fairness in Recommendation." WWW 2021
6. M. Hardt, E. Price, N. Srebro. "Equality of Opportunity in Supervised Learning." NeurIPS 2016

I have a working implementation ready with 22 passing tests and 0 regressions against the existing test suite. Happy to share early for feedback before opening the PR.

[miguelgfierro]

Love this idea @bajajahsaas, we actually have some responsible AI tests, and I think the tests for this should go under that category