ranked-overlap

ranked-overlap compares ordered lists using Rank-Biased Overlap (RBO). It is a production-ready version of the implementation introduced in my Towards Data Science article.

RBO is useful when rankings may have different lengths or contain different items, and when agreement near the top matters more than agreement near the bottom. Scores range from 0 (disjoint) to 1 (identical).

Installation

pip install ranked-overlap

Python API

from ranked_overlap import cumulative_weight, rbo

first = [1, 2, 3, 4, 5, 6, 7]
second = [1, 3, 2, 4, 5, 7, 6, 8]

score = rbo(first, second)
print(score)  # 0.8853713875

print(cumulative_weight(p=0.9, depth=10))  # 0.8555854467...

The persistence parameter p must lie strictly between 0 and 1. Lower values concentrate more weight at the beginning of each ranking. Rankings may be any iterable, but their items must be unique and hashable.

For compatibility with the original article, weightage_calculator(p, d) is available as an alias for cumulative_weight(p, depth).

Command line

ranked-overlap compare '["a", "b", "c"]' '["a", "c", "b"]' --p 0.9
ranked-overlap weight 10 --p 0.9

Development

python -m pip install -e '.[dev]'
ruff check .
mypy
pytest
python -m build
twine check dist/*

Citation

RBO was introduced by William Webber, Alistair Moffat, and Justin Zobel in “A Similarity Measure for Indefinite Rankings,” ACM Transactions on Information Systems, 2010. See CITATION.cff for citation metadata.

License

MIT © Krupesh Raikar