This repository can be used to reproduce results of "Word2vec applied to Recommendation: Hyperparameters Matter" by H. Caselles-Dupré, F. Lesaint and J. Royo-Letelier. The paper will be published on the 12th ACM Conference on Recommender Systems, Vancouver, Canada, 2nd-7th October 2018.

git clone https://github.com/deezer/w2v_reco_hyperparameters_matter.git

cd w2v_reco_hyperparameters_matter

docker build -t w2v_reco_hyperparameters_matter .

To reproduce results in Table 1: Next Event Prediction, line Fully optimized SGNS from paper:

docker run -ti --name=music_1_sgns w2v_reco_hyperparameters_matter:latest /bin/bash -c "python src/main.py --path_data='data/music_1.npy' --p2v=1 --window_size=3 --it=110 --sample=0.00001 --power_alpha=-0.5"

docker run -ti --name=music_2_sgns w2v_reco_hyperparameters_matter:latest /bin/bash -c "python src/main.py --path_data='data/music_2.npy' --p2v=1 --window_size=3 --it=130 --sample=0.00001 --power_alpha=-0.5"

docker run -ti --name=ecommerce_sgns w2v_reco_hyperparameters_matter:latest /bin/bash -c "python src/main.py --path_data='data/ecommerce_sessions.npy' --p2v=1 --window_size=3 --it=140 --sample=0.001 --power_alpha=1"

docker run -ti --name=kosarak_sgns w2v_reco_hyperparameters_matter:latest /bin/bash -c "python src/main.py --path_data='data/kosarak_sessions.npy' --p2v=1 --window_size=7 --it=150 --sample=0.00001 --power_alpha=-1"

To reproduce results in Table 2: NEP performance in cold-start regime, lines Fully optimized MetaProd2vec from paper:

docker run -ti --name=music_1_cs_mp2v_0 w2v_reco_hyperparameters_matter:latest /bin/bash -c "python src/main.py --path_data='data/music_1.npy' --p2v=0 --window_size=7 --it=90 --sample=0.0001 --power_alpha=-0.5 --cold_start=0"

docker run -ti --name=music_1_cs_mp2v_2 w2v_reco_hyperparameters_matter:latest /bin/bash -c "python src/main.py --path_data='data/music_1.npy' --p2v=0 --window_size=7 --it=90 --sample=0.0001 --power_alpha=-0.5 --cold_start=2"

docker run -ti --name=music_2_cs_mp2v_0 w2v_reco_hyperparameters_matter:latest /bin/bash -c "python src/main.py --path_data='data/music_2.npy' --p2v=0 --window_size=3 --it=150 --sample=0.0001 --power_alpha=-0.5 --cold_start=0"

docker run -ti --name=music_2_cs_mp2v_2 w2v_reco_hyperparameters_matter:latest /bin/bash -c "python src/main.py --path_data='data/music_2.npy' --p2v=0 --window_size=3 --it=150 --sample=0.0001 --power_alpha=-0.5 --cold_start=2"

Install gensim locally following https://radimrehurek.com/gensim/install.html.

git clone https://github.com/deezer/w2v_reco_hyperparameters_matter.git

cd w2v_reco_hyperparameters_matter

Replace the file gensim/models/word2vec.py in your local gensim installation folder by word2vec.py.

mkdir data

Copy .npy files from https://drive.google.com/drive/folders/1S-vneh5-egjzjNP7y1ChdevOjukQopqX to the data folder.

To reproduce results in Table 1: Next Event Prediction, line Fully optimized SGNS from paper:

python src/main.py --path_data='data/music_1.npy' --p2v=1 --window_size=3 --it=110 --sample=0.00001 --power_alpha=-0.5

python src/main.py --path_data='data/music_2.npy' --p2v=1 --window_size=3 --it=130 --sample=0.00001 --power_alpha=-0.5

python src/main.py --path_data='data/ecommerce_sessions.npy' --p2v=1 --window_size=3 --it=140 --sample=0.001 --power_alpha=1

python src/main.py --path_data='data/kosarak_sessions.npy' --p2v=1 --window_size=7 --it=150 --sample=0.00001 --power_alpha=-1

To reproduce results in Table 2: NEP performance in cold-start regime, lines Fully optimized MetaProd2vec from paper:

python src/main.py --path_data='data/music_1.npy' --p2v=0 --window_size=7 --it=90 --sample=0.0001 --power_alpha=-0.5 --cold_start=0

python src/main.py --path_data='data/music_1.npy' --p2v=0 --window_size=7 --it=90 --sample=0.0001 --power_alpha=-0.5 --cold_start=2

python src/main.py --path_data='data/music_2.npy' --p2v=0 --window_size=3 --it=150 --sample=0.0001 --power_alpha=-0.5 --cold_start=0

python src/main.py --path_data='data/music_2.npy' --p2v=0 --window_size=3 --it=150 --sample=0.0001 --power_alpha=-0.5 --cold_start=2