This repository provides a reference implementation for learning PiRank-based models as described in the paper:

PiRank: Learning to Rank via Differentiable Sorting
Robin Swezey, Aditya Grover, Bruno Charron and Stefano Ermon.
Paper: https://arxiv.org/abs/2012.06731

The codebase is implemented in Python 3.7. To install the necessary base requirements, run the following commands:

pip install -r requirements.txt

If you intend to use a GPU, modify requirements.txt to install tensorflow-gpu instead of tensorflow.

You will also need the NeuralSort implementation available here. Make sure it is added to your PYTHONPATH.

PiRank was tested on the two following datasets:

• The MSLR WEB30K data can be found at this address.
• The Yahoo! C14 dataset can be found at this address.

Additionally, the code is expected to work with any dataset stored in the standard LibSVM format used for LTR experiments.

There are two scripts for the code:

• pirank_simple.py implements a simple depth-1 PiRank loss (d=1). It is used in the experiments of sections 4.1 (benchmark evaluation on MSLR-WEB30K and Yahoo! C14 datasets), 4.2.1 (effect of temperature parameter), and 4.2.2 (effect of training list size).

• pirank_deep.py implements the deeper PiRank losses (d>=1). It is used for the experiments of section 4.2.3 and comes with a convenient synthetic data generator as well as more tuning options.

Options are handled by Sacred (see Examples section below).

PiRank-related:

NeuralSort-related:

TensorFlow-Ranking and architecture-related:

Training-related:

Sacred:

In addition, you can use regular parameters from Sacred (such as -m for logging the experiment to MongoDB).

Run the benchmark experiment of section 4.1 with PiRank simple loss on MSLR-WEB30K

cd pirank
python3 pirank_simple.py with loss_fn=pirank_simple_loss \
    ndcg_k=10 \
    tau=5 \
    list_size=80 \
    hidden_layers=256,relu,256,relu,128,relu,64,relu \
    train_path=/data/MSLR-WEB30K/Fold1/train.txt \
    vali_path=/data/MSLR-WEB30K/Fold1/vali.txt \
    test_path=/data/MSLR-WEB30K/Fold1/test.txt \
    num_features=136 \
    optimizer=Adam \
    learning_rate=0.00001 \
    num_epochs=100 \
    batch_size=16 \
    model_dir=/tmp/model

Run the benchmark experiment of section 4.1 with PiRank simple loss on Yahoo! C14

cd pirank
python3 pirank_simple.py with loss_fn=pirank_simple_loss \
    ndcg_k=10 \
    tau=5 \
    list_size=80 \
    hidden_layers=256,relu,256,relu,128,relu,64,relu \
    train_path=/data/YAHOO/set1.train.txt \
    vali_path=/data/YAHOO/set1.valid.txt \
    test_path=/data/YAHOO/set1.test.txt \
    num_features=700 \
    optimizer=Adam \
    learning_rate=0.00001 \
    num_epochs=100 \
    batch_size=16 \
    model_dir=/tmp/model

Run the benchmark experiment of section 4.1 with classic LambdaRank on MSLR-WEB30K

cd pirank
python3 pirank_simple.py with loss_fn=lambda_rank_loss \
    ndcg_k=10 \
    tau=5 \
    list_size=80 \
    hidden_layers=256,relu,256,relu,128,relu,64,relu \
    train_path=/data/MSLR-WEB30K/Fold1/train.txt \
    vali_path=/data/MSLR-WEB30K/Fold1/vali.txt \
    test_path=/data/MSLR-WEB30K/Fold1/test.txt \
    num_features=136 \
    optimizer=Adam \
    learning_rate=0.00001 \
    num_epochs=100 \
    batch_size=16 \
    model_dir=/tmp/model

Run the scaling ablation experiment of section 4.2.3 using synthetic data generation (d=2)

cd pirank
python3 pirank_deep.py with loss_fn=pirank_deep_loss \
    ndcg_k=10 \
    ste=True \
    merge_block_size=100 \
    tau=5 \
    taustar=1e-10 \
    tau_scheme=square \
    data_generator=synthetic_data_generator \
    actual_list_size=1000 \
    list_size=1000 \
    vali_list_size=1000 \
    test_list_size=1000 \
    full_loss=False \
    train_path=0 \
    vali_path=1 \
    test_path=2 \
    num_queries=1000 \
    num_features=25 \
    num_query_features=5 \
    hidden_layers=256,relu,256,relu,128,relu,128,relu,64,relu,64,relu \
    optimizer=Adam \
    learning_rate=0.00001 \
    num_epochs=100 \
    batch_size=16

If you need help, reach out to Robin Swezey or raise an issue.

If you find PiRank useful in your research, please consider citing the following paper:

@inproceedings{
swezey2020pirank,
title={PiRank: Learning to Rank via Differentiable Sorting},
author={Robin Swezey and Aditya Grover and Bruno Charron and Stefano Ermon},
year={2020},
url={},
}