Unbiased Learning to Rank Algorithms (ULTRA)

🔥News: A TensorFlow version of this package can be found in ULTRA.

This is an Unbiased Learning To Rank Algorithms (ULTRA) toolbox, which provides a codebase for experiments and research on learning to rank with human annotated or noisy labels. With the unified data processing pipeline, ULTRA supports multiple unbiased learning-to-rank algorithms, online learning-to-rank algorithms, neural learning-to-rank models, as well as different methods to use and simulate noisy labels (e.g., clicks) to train and test different algorithms/ranking models. A user-friendly documentation can be found here.

Get Started

Create virtual environment (optional):

pip install --user virtualenv
~/.local/bin/virtualenv -p python3 ./venv
source venv/bin/activate

Install ULTRA from the source:

git clone https://github.com/ULTR-Community/ULTRA_pytorch.git
cd ULTRA
make init

Run toy example:

bash example/toy/offline_exp_pipeline.sh

Structure

Input Layers

ClickSimulationFeed: this is the input layer that generate synthetic clicks on fixed ranked lists to feed the learning algorithm.
DeterministicOnlineSimulationFeed: this is the input layer that first create ranked lists by sorting documents according to the current ranking model, and then generate synthetic clicks on the lists to feed the learning algorithm. It can do result interleaving if required by the learning algorithm.
StochasticOnlineSimulationFeed: this is the input layer that first create ranked lists by sampling documents based on their scores in the current ranking model and the Plackett-Luce distribution, and then generate synthetic clicks on the lists to feed the learning algorithm. It can do result interleaving if required by the learning algorithm.
DirectLabelFeed: this is the input layer that directly feed the labels of each documents (e.g., the true relevance labels or raw click logs) to the learning algorithm.

Learning Algorithms

NA: this model is an implementation of the naive algorithm that directly train models with input labels (e.g., clicks).
DLA: this is an implementation of the Dual Learning Algorithm in Unbiased Learning to Rank with Unbiased Propensity Estimation.
IPW: this model is an implementation of the Inverse Propensity Weighting algorithms in Learning to Rank with Selection Bias in Personal Search and Unbiased Learning-to-Rank with Biased Feedback
REM: this model is an implementation of the regression-based EM algorithm in Position bias estimation for unbiased learning to rank in personal search
PD: this model is an implementation of the pairwise debiasing algorithm in Unbiased LambdaMART: An Unbiased Pairwise Learning-to-Rank Algorithm.
DBGD: this model is an implementation of the Dual Bandit Gradient Descent algorithm in Interactively optimizing information retrieval systems as a dueling bandits problem
MGD: this model is an implementation of the Multileave Gradient Descent in Multileave Gradient Descent for Fast Online Learning to Rank
NSGD: this model is an implementation of the Null Space Gradient Descent algorithm in Efficient Exploration of Gradient Space for Online Learning to Rank
PDGD: this model is an implementation of the Pairwise Differentiable Gradient Descent algorithm in Differentiable unbiased online learning to rank

Ranking Models

Linear: this is a linear ranking algorithm that compute ranking scores with a linear function.
DNN: this is neural ranking algorithm that compute ranking scores with a multi-layer perceptron network (with non-linear activation functions).
DLCM: this is an implementation of the Deep Listwise Context Model in Learning a Deep Listwise Context Model for Ranking Refinement (TODO).
GSF: this is an implementation of the Groupwise Scoring Function in Learning Groupwise Multivariate Scoring Functions Using Deep Neural Networks (TODO).
SetRank: this is an implementation of the SetRank model in SetRank: Learning a Permutation-Invariant Ranking Model for Information Retrieval (TODO).

Supported Evaluation Metrics

MRR: the Mean Reciprocal Rank.
ERR: the Expected Reciprocal Rank from Expected reciprocal rank for graded relevance.
ARP: the Average Relevance Position.
NDCG: the Normalized Discounted Cumulative Gain.
DCG: the Discounted Cumulative Gain.
Precision: the Precision.
MAP: the Mean Average Precision.
Ordered_Pair_Accuracy: the percentage of correctedly ordered pair.

Click Simulation Example

Create click models for click simulations

python ultra/utils/click_models.py pbm 0.1 1 4 1.0 example/ClickModel

* The output is a json file containing the click mode that could be used for click simulation. More details could be found in the code.

(Optional) Estimate examination propensity with result randomization

python ultra/utils/propensity_estimator.py example/ClickModel/pbm_0.1_1.0_4_1.0.json <DATA_DIR> example/PropensityEstimator/

* The output is a json file containing the estimated examination propensity (used for IPW). DATA_DIR is the directory for the prepared data created by ./libsvm_tools/prepare_exp_data_with_svmrank.py. More details could be found in the code.

Citation

If you use ULTRA in your research, please use the following BibTex entry.

@misc{tran2021ultra,
      title={ULTRA: An Unbiased Learning To Rank Algorithm Toolbox}, 
      author={Anh Tran and Tao Yang and Qingyao Ai},
      year={2021},
      eprint={2108.05073},
      archivePrefix={arXiv},
      primaryClass={cs.IR}
}

@article{10.1145/3439861,
author = {Ai, Qingyao and Yang, Tao and Wang, Huazheng and Mao, Jiaxin},
title = {Unbiased Learning to Rank: Online or Offline?},
year = {2021},
issue_date = {February 2021},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
volume = {39},
number = {2},
issn = {1046-8188},
url = {https://doi.org/10.1145/3439861},
doi = {10.1145/3439861},
journal = {ACM Trans. Inf. Syst.},
month = feb,
articleno = {21},
numpages = {29},
keywords = {unbiased learning, online learning, Learning to rank}
}