SSLRec does not seem to be able to effectively reproduce some of the baseline results about sslrec HOT 1 CLOSED

Hi!

Thanks for your interest! The code for sampling positive/negative items can be found in the data_utils/datasets_general_cf.py file, as shown below:

class PairwiseTrnData(data.Dataset):
    def __init__(self, coomat):
    self.rows = coomat.row
    self.cols = coomat.col
    self.dokmat = coomat.todok()
    self.negs = np.zeros(len(self.rows)).astype(np.int32)
    
    def sample_negs(self):
    for i in range(len(self.rows)):
	u = self.rows[i]
	while True:
            iNeg = np.random.randint(configs['data']['item_num'])
            if (u, iNeg) not in self.dokmat:
                break
	self.negs[i] = iNeg
    
    def __len__(self):
    return len(self.rows)
    
    def __getitem__(self, idx):
    return self.rows[idx], self.cols[idx], self.negs[idx]

While it may differ from the sampling strategy used in LightGCN-PyTorch, which employs uniform sampling by first sampling users and then sampling both positive and negative items, SSLRec follows the distribution of the original interactions. This means that we sample negative items for each interaction during training.

Additionally, during training, we employ early stopping to avoid manually tuning the number of training epochs for all methods. This approach has proven effective for the currently implemented methods. However, when dealing with new datasets, the hyperparameters for early stopping (such as the patience) may vary to achieve optimal performance for different methods.

Moreover, several model-agnostic sub-details (models/model_utils.py, models/loss_utils.py), such as InfoNCE loss and regularization loss, are uniformly implemented in our SSLRec for all methods. However, it is important to note that the implementation may differ from other repositories. Each repository has its own unique design and implementation approach. Nonetheless, in all our current experiments, SimGCL consistently outperforms LightGCN in various scenarios and datasets.

Lastly, thank you for testing our framework and for your valuable questions. If you have any ideas to improve SSLRec, please don't hesitate to submit a pull request (PR). We appreciate your contributions to enhance the SSLRec framework : )

Best regards,
Xubin

from sslrec.