The repository implement the Bayesian Personalized Ranking using pyTorch (https://arxiv.org/pdf/1205.2618)
Other repositories also implement this model, but the evaluation takes longer time.
So, I implement this model using pyTorch with GPU acceleration for evaluation.
Implementation detail will be explained in the following section.

You have to install the following package before executing this code.

• python==3.6
• pytorch==1.0.0
• numpy==1.15.4
• pandas==0.23.4

You can install these package by executing the following command or through anaconda.

pip install -r requirements.txt

This code support the movielens 1m data and movielens 20m data. You can get the dataset from the following list.

• MovieLens 1M.
• MovieLens 20M.

After downloading that file, unzip it.
We call the path for unzipped directory $data_dir.

For basic usage, execute following command line to preprocess the data. It randomly split the whole dataset into two parts, training data and test data.

python preprocess.py --dataset ml-1m --data_dir $data_dir --output_data preprocess/ml-1m.pickle
python preprocess.py --dataset ml-20m --data_dir $data_dir --output_data preprocess/ml-20m.pickle

If you want to split training data and test data with time order, then execute the following command line. This code sorts the item list for each user using time order. After that, it splits the whole data into two parts, training data and test data. First 80% of the item list will become the training data and the last 20% of the item list will become test data.

python preprocess.py --dataset ml-1m --output_data preprocess/ml-1m.pickle --time_order
python preprocess.py --dataset ml-20m --output_data preprocess/ml-20m.pickle --time_order

Help message will give you more detail description for arguments.

python preprocess.py --help

Now, for real show, let's train MF model using BPR-OPT loss. You can execute the following command to train MF model using BPR-OPT.

python train.py --data preprocess/ml-1m.pickle
python train.py --data preprocess/ml-20m.pickle

Help message will give you more detail description for arguments. You can train MF model with different hyperparameter.

python train.py --help

• I didn't use regularization coefficient for each embedding matrix. If you want to tune each coefficient, then you need to add regularization term inside the loss value.

The evaluation benchmark for movielens 1m is the following table. I think more tuning will get better result, but this value is reasonably around the statistic.

More information will get from the result directory.

Q. Loss doesn't decrease, why? A. I print the smoothing loss and the initial smoothing loss is 0, which cause this problem. Waiting just a moment, then loss will go down.

Q. Loss converge to 0.6931. A. Because weight decay is so strong that model cannot learn from dataset. Decrease the weight decay factor.

If you have any problem during simulating this code, open issue or contact me by sending email to [email protected]