ranklab 0.1.1

Learning to Rank: differentiable ranking, LTR losses (RankNet, LambdaRank, ApproxNDCG, ListNet, ListMLE), and IR evaluation metrics.

ranklab

Differentiable ranking, learning-to-rank losses, and IR evaluation metrics, backed by a Rust implementation for low overhead per call.

Python bindings for the rankit Rust crate.

Install

pip install ranklab

Quick start

import ranklab

# Differentiable soft ranking
scores = [5.0, 1.0, 2.0, 4.0, 3.0]
ranks = ranklab.soft_rank(scores, temperature=1.0)

# RankNet pairwise loss
predictions = [0.8, 0.3, 0.6]
relevance = [2.0, 0.0, 1.0]
loss = ranklab.ranknet_loss(predictions, relevance)

# NDCG evaluation
ranked = [("doc1", 0.9), ("doc2", 0.8), ("doc3", 0.7)]
qrels = {"doc1": 2, "doc2": 1, "doc3": 0}
score = ranklab.ndcg(ranked, qrels, k=3)

API

Name	Description
soft_rank	Soft ranking via optimal-transport relaxation
soft_rank_neural_sort	Soft ranking via NeuralSort relaxation
soft_rank_sigmoid	Soft ranking via sigmoid pairwise comparisons
soft_rank_smooth_i	Soft ranking via smooth indicator functions
differentiable_topk	Differentiable top-k selection, returns (values, indicators)
ranknet_loss	RankNet pairwise cross-entropy loss
approx_ndcg	Differentiable NDCG approximation via softmax
lambda_loss	LambdaLoss ranking loss
listnet_loss	ListNet top-1 probability loss (KL divergence)
listmle_loss	ListMLE likelihood loss for permutation learning
compute_lambdarank_gradients	Per-document LambdaRank gradients
compute_ranking_svm_gradients	Per-document RankingSVM gradients
ndcg	Normalized Discounted Cumulative Gain at k
map_score	Mean Average Precision
mrr	Mean Reciprocal Rank
precision_at_k	Precision at depth k
recall_at_k	Recall at depth k
err_at_k	Expected Reciprocal Rank at depth k
rbp_at_k	Rank-Biased Precision at depth k
f_measure_at_k	F-measure (harmonic mean of P and R) at depth k
success_at_k	Whether any relevant doc appears in top-k
r_precision	Precision at R (R = number of relevant docs)
average_precision	Average Precision for binary relevance
paired_t_test	Paired t-test on two score sets, returns dict
confidence_interval	Confidence interval for a score set
cohens_d	Cohen's d effect size
load_trec_run	Parse a TREC run file into {query_id: [(doc_id, score)]}
load_qrels	Parse a TREC qrels file into {query_id: {doc_id: rel}}
evaluate_batch	Batch evaluation across multiple queries
evaluate_trec	Evaluate TREC run+qrels files with batch metrics

numpy support

Scoring and loss functions accept numpy arrays or Python lists. Gradient functions return numpy float32 arrays. Eval metrics take list[tuple[str, float]] for ranked results and dict[str, int] for relevance judgments.

License

MIT OR Apache-2.0