BPR_PyTorch

Bayesian Personalized Ranking with PyTorch.

Environment

• Python: 3.6
• PyTorch: 1.5.1
• CUDA: 10.1
• Ubuntu: 18.04

Dataset

The Movielens 1M Dataset is used. The rating data is included in data/ml-1m.

Run the Codes

$ python BPR_PyTorch/main.py

Details

For each user, the latest and the second latest rating are used as test and validation, respectively. The remaining ratings are used as training. The hyperparameters (batch_size, lr, latent_dim, l2_reg) are tuned by using the valudation data in terms of nDCG. See config.ini about the range of each hyperparameter.

Although the original ratings range 1 to 5, all of them are used as positive data. Items that are not consumed by a user are used as negative data for the user.

By running the code, hyperparameters are automatically tuned. After the training process, the best hyperparameters and HR/nDCG computed by using the test data are displayed.

Given a specific combination of hyperparameters, the corresponding training results are saved in data/train_result/<hyperparameter combination> (e.g., data/train_result/batch_size_512-lr_0.005-latent_dim_8-l2_reg_1e-07-epoch_3-n_negative_4-top_k_10). In the directory, a model file (model.pth) and a json file (epoch_data.json) that describes information for each epoch are generated. The json file can be described as follows (epoch=3).

[
    {
        "epoch": 0,
        "loss": 4994338.738952637,
        "HR": 0.0728476821192053,
        "NDCG": 0.03107992452383637
    },
    {
        "epoch": 1,
        "loss": 3138874.3325195312,
        "HR": 0.14966887417218544,
        "NDCG": 0.06523117735080695
    },
    {
        "epoch": 2,
        "loss": 1708930.3295593262,
        "HR": 0.4064569536423841,
        "NDCG": 0.21713509086237084
    }
]