ranx 0.2.13

ranx: A Blazing Fast Python Library for Ranking Evaluation and Comparison

🔥 News

• 📌 [July 27, 2022] ranx will be featured in CIKM 2022, the 31st ACM International Conference on Information and Knowledge Management!

• [August 29, 2022] ranx 0.2.9 is out.
  Filetypes are now automatically inferred from file extensions (.json → json, .trec → trec, .txt → trec). Default behavior can be overridden with the kind parameter (this should allow for backward compatibility).
  Two-sided Paired Student's t-Test is now the default statistical test used when calling compare (it is much faster than Fisher's and they usually agree).
  Loading / saving Qrels and Run from / to json files is now much faster thanks to orjson.

• [June 29, 2022] Added support for Tukey's HSD Test.

• [June 28, 2022] Added support for Bpref and Rank-biased Precision (RBP) metrics.

• [June 9, 2022] Added support for 25 fusion algorithms, six normalization strategies, and an automatic fusion optimization functionality in v.0.2.
  Check out the official documentation and Jupyter Notebook for further details on fusion and normalization.

⚡️ Introduction

ranx is a library of fast ranking evaluation metrics implemented in Python, leveraging Numba for high-speed vector operations and automatic parallelization. It offers a user-friendly interface to evaluate and compare Information Retrieval and Recommender Systems. ranx allows you to perform statistical tests and export LaTeX tables for your scientific publications. Moreover, ranx provides several fusion algorithms and normalization strategies, and an automatic fusion optimization functionality. ranx was featured in ECIR 2022, the 44th European Conference on Information Retrieval.

If you use ranx to evaluate results or conducting experiments involving fusion for your scientific publication, please consider citing it.

For a quick overview, follow the Usage section.

For a in-depth overview, follow the Examples section.

✨ Features

Metrics

• Hits
• Hit Rate
• Precision
• Recall
• F1
• r-Precision
• Bpref
• Rank-biased Precision (RBP)
• Mean Reciprocal Rank (MRR)
• Mean Average Precision (MAP)
• Normalized Discounted Cumulative Gain (NDCG)

The metrics have been tested against TREC Eval for correctness.

Statistical Tests

• Fisher's Randomization Test
• Paired Student's t-Test (default)
• Tukey's HSD Test

Please, refer to Smucker et al., Carterette, and Fuhr for additional information on statistical tests for Information Retrieval.

Off-the-shelf Qrels

You can load qrels from ir-datasets as simply as:

qrels = Qrels.from_ir_datasets("msmarco-document/dev")

A full list of the available qrels is provided here.

Fusion Algorithms

Name	Name	Name	Name	Name
CombMIN	CombMNZ	RRF	MAPFuse	BordaFuse
CombMED	CombGMNZ	RBC	PosFuse	Weighted BordaFuse
CombANZ	ISR	WMNZ	ProbFuse	Condorcet
CombMAX	Log_ISR	Mixed	SegFuse	Weighted Condorcet
CombSUM	LogN_ISR	BayesFuse	SlideFuse	Weighted Sum

Please, refer to the documentation for further details.

Normalization Strategies

• Min-Max Norm
• Max Norm
• Sum Norm
• ZMUV Norm
• Rank Norm
• Borda Norm

Please, refer to the documentation for further details.

🔌 Installation

pip install ranx

💡 Usage

Create Qrels and Run

from ranx import Qrels, Run

qrels_dict = { "q_1": { "d_12": 5, "d_25": 3 },
               "q_2": { "d_11": 6, "d_22": 1 } }

run_dict = { "q_1": { "d_12": 0.9, "d_23": 0.8, "d_25": 0.7,
                      "d_36": 0.6, "d_32": 0.5, "d_35": 0.4  },
             "q_2": { "d_12": 0.9, "d_11": 0.8, "d_25": 0.7,
                      "d_36": 0.6, "d_22": 0.5, "d_35": 0.4  } }

qrels = Qrels(qrels_dict)
run = Run(run_dict)

Evaluate

from ranx import evaluate

# Compute score for a single metric
evaluate(qrels, run, "ndcg@5")
>>> 0.7861

# Compute scores for multiple metrics at once
evaluate(qrels, run, ["map@5", "mrr"])
>>> {"map@5": 0.6416, "mrr": 0.75}

Compare

from ranx import compare

# Compare different runs and perform Two-sided Paired Student's t-Test
report = compare(
    qrels=qrels,
    runs=[run_1, run_2, run_3, run_4, run_5],
    metrics=["map@100", "mrr@100", "ndcg@10"],
    max_p=0.01  # P-value threshold
)

Output:

print(report)

#    Model    MAP@100    MRR@100    NDCG@10
---  -------  --------   --------   ---------
a    model_1  0.320ᵇ     0.320ᵇ     0.368ᵇᶜ
b    model_2  0.233      0.234      0.239
c    model_3  0.308ᵇ     0.309ᵇ     0.330ᵇ
d    model_4  0.366ᵃᵇᶜ   0.367ᵃᵇᶜ   0.408ᵃᵇᶜ
e    model_5  0.405ᵃᵇᶜᵈ  0.406ᵃᵇᶜᵈ  0.451ᵃᵇᶜᵈ

Fusion

from ranx import fuse, optimize_fusion

best_params = optimize_fusion(
    qrels=train_qrels,
    runs=[train_run_1, train_run_2, train_run_3],
    norm="min-max",     # The norm. to apply before fusion
    method="wsum",      # The fusion algorithm to use (Weighted Sum)
    metric="ndcg@100",  # The metric to maximize
)

combined_test_run = fuse(
    runs=[test_run_1, test_run_2, test_run_3],  
    norm="min-max",       
    method="wsum",        
    params=best_params,
)

📖 Examples

Name	Link
Overview
Qrels and Run
Evaluation
Comparison and Report
Fusion

📚 Documentation

Browse the documentation for more details and examples.

🎓 Citation

If you use ranx to evaluate results for your scientific publication, please consider citing it:

@inproceedings{bassani2022ranx,
  author    = {Elias Bassani},
  title     = {ranx: {A} Blazing-Fast Python Library for Ranking Evaluation and Comparison},
  booktitle = {{ECIR} {(2)}},
  series    = {Lecture Notes in Computer Science},
  volume    = {13186},
  pages     = {259--264},
  publisher = {Springer},
  year      = {2022}
}

🎁 Feature Requests

Would you like to see other features implemented? Please, open a feature request.

🤘 Want to contribute?

Would you like to contribute? Please, drop me an e-mail.

📄 License

ranx is an open-sourced software licensed under the MIT license.