CFNADE-keras

Paper Description

• Paper: "A Neural Autoregressive Approach to Collaborative Filtering" LINK
• Unofficial Slide: LINK
• Keras implementation of CF-NADE.
• The CFNADE is the model applying NADE model to the collaborative filtering problem. NADE is the modified version of the auto-encoder to estimate probabilistic distribution.

Limitation

• This repository only works with Movielens 1M data.
• This repository force you to use "tensorflow" backend.

How to run

• I used python3.

git clone git@github.com:JoonyoungYi/CFNADE-keras.git
cd CFNADE-keras
virtualenv .venv -p python3
. .venv/bin/activate
pip install -r requirements.txt
python data_prep.py
python run.py

• I setup my environment by this docker image.

Performance

Setting 1.

• default setting.

--hidden_dim=250

• 1st Attempt.

training set RMSE for epoch 29 is 0.858837
validation set RMSE for epoch 29 is 0.872606
Testing...
test set RMSE is 0.820515

• 2nd Attempt.

training set RMSE for epoch 29 is 0.882864
validation set RMSE for epoch 29 is 0.874668
val_nade_loss_loss: 0.0000e+00
Testing...
test set RMSE is 0.823059

• 3rd Attempt.

training set RMSE for epoch 29 is 0.868767
validation set RMSE for epoch 29 is 0.874799
Testing...
test set RMSE is 0.824332

Setting 2.

--hidden_dim=500

• 1st Attempt.

training set RMSE for epoch 29 is 0.844765
validation set RMSE for epoch 29 is 0.873419
Testing...
test set RMSE is 0.793768

Repository History

• keras implementation of CF-NADE based on the implementation of Ian09/CF-NADE
• Forked from AlexGidiotis/keras-CF-NADE