Multi-objective Learning to Rank through Multi-objective Optimization with LightGBM

Introduction

MO-LightGBM is a gradient boosting framework based on decision tree algorithms, used for multi-label learning to rank tasks. It is based on LightGBM.

Requirements

• Python 3.8 or higher
• matplotlib
• seaborn

Installation

1. Install blas and lapack,

   1. In mac:

      $brew install openblas
      $brew install lapack
      

   2. In Debian-based systems like Ubuntu:

      $sudo apt install libblas-dev liblapack-dev
      

   3. In Red Hat systems like CentOS and Fedora:

      $sudo yum install lapack-devel blas-devel
      

2. From the scs root, build the package, install cmake if needed

   cd LightGBM/external_libs/scs/
   rm -r build
   mkdir build
   cd build
   cmake ..
   make -j$(nproc)
   cd ../../../../
   

3. From the ecos root, build the package

   cd LightGBM/external_libs/ecos.2.0.8/
   rm -r build
   mkdir build
   cd build
   cmake ..
   make -j$(nproc)
   cd ../../../../
   

4. Now build LightGBM

   cd LightGBM
   rm -r build
   mkdir build
   cd build
   cmake ..
   make -j$(nproc)
   cd ../../
   

It is fully tested with Red Hat systems.

Usage

Assumptions about Dataset

1. The dataset should be in csv or tsv format, not in the libsvm format.
2. The csv/tsv file should have header that names each column

Process for bi-objective experiments

1. Prepare the <dataset_name>_config.yml

2. Generate configure files for baseline experiment

   python baselines.py <dataset_name>_config.yml

3. Go to the baseline_results folder and run the experiments

   cd results_${dataset}/lambdarank/baseline_results
   sh run_experiment.sh
   

4. Generate configure files for biobjective experiments for preference based methods

   python biobjectives.py <dataset_name>_config.yml

5. Go to the biobjective_results folder, and, either run all the experiments or run for individual combinator

   cd results_${dataset}/lambdarank/biobjective_results
   sh run_experiment.sh   # will run for the all combinators
   sh epo_search_run_experiment.sh     # will run only for epo_search  
   

6. Generate configure files for biobjective experiments for constraints based methods

   python biobjectives_ec.py <dataset_name>_config.yml

7. Go to the biobjective_results folder, and, either run all the experiments or run for individual combinator

   cd results_${dataset}/lambdarank/biobjective_results
   sh ec_run_experiment.sh   # will run for the all combinators
   

8. [optional] Run plotting script

   1. Plot bound and ray in one figure

      python plot_biobjectives.py <dataset_name>_config.yml
      

Example for bi-objective experiment

dataset=istella

root=/home/user/MO-LightGBM/

folder=${root}/example/

cd ${folder}

python baselines.py ${dataset}_config.yml

cd results_${dataset}/lambdarank/baseline_results

sh run_experiment.sh

cd ${folder}


python biobjectives.py ${dataset}_config.yml

cd results_${dataset}/lambdarank/biobjective_results

sh run_experiment.sh 

cd ${folder}


python biobjectives_ec.py ${dataset}_config.yml

cd results_${dataset}/lambdarank/biobjective_results

sh ec_run_experiment.sh 

cd ${folder}


python plot_biobjectives.py ${dataset}_config.yml

Citation

If you use this work, or otherwise found our work valuable, please consider citing the papers:

Multi-Label Learning to Rank through Multi-Objective Optimization

@article{chaoshengmo-lightgbm2025,
  title={MO-LightGBM: A Library for Multi-objective Learning to Rank with LightGBM},
  author={Dong, Chaosheng and Momma, Michinari},
  year={2025}
}

@inproceedings{mahapatra2023multi,
  title={Multi-label learning to rank through multi-objective optimization},
  author={Mahapatra, Debabrata and Dong, Chaosheng and Chen, Yetian and Momma, Michinari},
  booktitle={Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
  year={2023}
}

Security

See CONTRIBUTING for more information.

License

This project is licensed under the Apache-2.0 License.