Scalable Ranking under Constraints

The repository contains code to reproduce results in the working paper Yegor Tkachenko, Wassim Dhaouadi, and Kamel Jedidi. Scaling up Ranking under Constraints for Live Recommendations by Replacing Optimization with Prediction (2022).

Here is the bibtex citation format:

@article{scalableranking2022,
	author  = {Tkachenko, Yegor and Dhaouadi, Wassim and Jedidi, Kamel},
	title   = {{Scaling up Ranking under Constraints for Live Recommendations by Replacing Optimization with Prediction}},
	journal = {arXiv preprint arXiv:2202.07088},
	year    = {2022}
}

The paper describes a methodology for real-time optimal re-ranking of content candidates based on multiple objectives - as the final stage in a recommender system.

./code/environment.yml is a conda environment file you can use to set up the programming environment.

You will need to download GroupLens 25M and Yow News data sets such that you get the following directory structure.

• data
  • yow-dataset
    • yow_userstudy_raw.xls
  • ml-25m
    • ratings.csv
    • movies.csv
    • genome-scores.csv
    • ...
• code

You will then run the python code files from the code directory in the following order to generate the data, run experiments, and visualize the results (run.sh file is provided for convenience):

• prepare_data_mf_movies.py
• prepare_data_mf_news.py
• ech_movies_50.py
• ech_movies_500.py
• ech_movies_1000.py
• ech_news_50.py
• ech_news_500.py
• ech_news_1000.py
• summarize_results.py

Note:

• We use CBC solver, via cvxpy interface, as can be seen in core_functions_unbalanced.py. The solver needs to be installed separately.
• Variable EPSILON in code refers to quantity in the paper.

Copyright (C) 2022 Yegor Tkachenko, Wassim Dhaouadi, Kamel Jedidi