ryubright/BPR

PyTorch Implementation of "BPR: Bayesian Personalized Ranking from Implicit Feedback"

BPR

PyTorch Implementation of "BPR: Bayesian Personalized Ranking from Implicit Feedback"

I used the dataset from this repo MCRec and ml-100k.

1. BPR-OPT

BPR-OPT is an optimization technique for personalized ranking.

def bpr_loss(pos_scores, neg_scores):
    loss = -(pos_scores - neg_scores).sigmoid().log().sum()
    return loss

2. Requirements

numpy==1.24.1
pandas==1.5.2
scikit_learn==1.2.1
scikit_surprise==1.1.3
surprise==0.1
torch==1.13.1
torcheval==0.0.6
tqdm==4.64.1

3. Example run

• set config.json

{
  "seed": 417,
  "n_neg": 4,
  "batch_size": 4096,
  "n_factor": 32,
  "learning_rate": 0.001,
  "weight_decay": 0,
  "epochs": 50,
  "top_k": 10,
  "test_file_path": "./data/ml-100k.test.negative",
  "test_neg_n": 100,
  "save_dir": "./save_models"
}

• run python code

python train.py --config [config.json file path]