Master Thesis: Popularity Bias and Fairness in Recommendation

Acknowledgment

• The implementation of ALS is modified from (https://github.com/tushushu/imylu/tree/master/imylu/recommend)
• The implementation of VAE-CF is modified from (https://github.com/dawenl/vaecf)
• The implementation of Long-tail GAN is modified from (https://github.com/CrowdDynamicsLab/NCF-GAN)

Final Dissertation

The final version of dissertation can be found here, which all the explanation and figures are presented to show own work.

Experiments

Setup Environment

First of all, install all the required packages using

conda install numpy=1.16.0 tensorflow-gpu=1.14.0 scikit-learn=0.23.1 seaborn=0.11.1 matplotlib=3.3.2 pandas=1.1.3 bottleneck=1.3.2 psutil=5.8.0 scipy=1.5.2 configparser=5.0.2

Dataset Processing

Before running the experiments, please check if the dataset exists in the first level directory raw_data. As long as the dataset exist, the DataConvert_dat_to_csv.ipynb can be used for dataset converting. Then, process_rawdata.ipynb is used to split dataset(with holdout ratings) by first approach and generated different group of users. The second approach is implemented together with VAE_CF(holdout users).ipynb

ALS

In order to run the experiments for ALS-MF, please use the ALS.ipynb file under the first level directory ALS. Test section is alse include in this file.

VAE-CF

In order to run the experiments for VAE-CF, please use the VAE_CF(holdout_ratings).ipynb and VAE_CF(holdout_users).ipynb. Test section is alse include in this file.

Long-tail GAN

In order to run the experiments for long-tail GAN, please use the following commend lines to run the experiments. For training:

	cd long-tail_GAN/Codes
	python train.py ../Dataset/ml_1m_holdout_user

For testing:

	cd long-tail_GAN/Codes
	python test.py ../Dataset/ml_1m_holdout_user