Bernoulli Negative Sampling in DICEE models?

[zg009]

Hello, and thank you for the library.

I am coming from PyKEEN due to newer models in this library and I wanted to know if there is an API for integrating negative sampling choices in the arguments, or if negative sampling strategies are already implemented in the models during training. Thank you

[Demirrr]

Hi @zg009,

We have "AllvsAll", "KvsAll", "1vsAll", "NegSample", "1vsSample", "KvsSample" as techniques to generate negative examples.
Bernoulli Negative sampling can be implemented by creating a new dataset class in dataset_classes.py

Please take a look the dataset classes implemented therein, particularly

dice-embeddings/dicee/dataset_classes.py

Line 641 in cce388c

class TriplePredictionDataset(torch.utils.data.Dataset):

You can very well the corruption logic based on the relation in the collate_fn.

Please implement the dataset class and make a PR If it is not too much of work for you, otherwise I can take care of the work around Dec.

[zg009]

Hello,
Thank you for your prompt response.

I looked through the examples and I could see where "MultiLabelDataset", "TriplePredictionDataset", "NegSampleDataset", "OnevsSample" and "CVDataModule" use negative sampling ratio. I do not see negative sampling in "KvsAll" - I am guessing this is because there is no need to generate negative samples in a KvsAll training?

I do not need to re-implement Bernoulli sampling as I already have some code which performs it. I do not want to add the negative sample to DICEE training set and have it also generate negative samples if some already exist in my dataset.