Winning Solution of PCIC 2021: Causal Inference and Recommendation(link)

This repository provides our winning solution for the PCIC 2021: Causal Inference and Recommendation.

If you have any questions, please feel free to contact by issues or yitianartsky@gmail.com.

Introduction

• We proposed and design a debias model which leverage the rating data for more accurate prediction.
• The file "NNEnsembleSubmit.py" generates two part results: one from the basic debias model framework, the other was achieved by weighted training based on inverse propensity score.
• Users can get the intermediate results of the debias model: the 'dotProduct', 'userBias', 'itemBias' through the file "dataPreprocessMovie.py".
• For those data which include "userid, tagid, rating, aveMovieRate" we retrain the model, which generate more accurate prediction results due to richer features: the.

Reproduce result

* run "python3 NNEnsembleSubmit.py" to generate data file "submit20210826.csv".
* run "dataPreprocessMovie.py" to generate  files "testDt.csv, validDt.csv, totalDt.csv"
* run "Rscript stackModeling.R" to generate final submit file "revision20210825.csv".