Awesome

M-Mix: Generating Hard Negatives via Multi-sample Mixing for Contrastive Learning

Code of SIGKDD 22 paper "M-Mix: Generating Hard Negatives via Multi-sample Mixing for Contrastive Learning"

This paper proposes to mix multiple samples in one mini-batch to generate hard negative pairs.

To pre-train the encoder on CIFAR-10 and CIFAR-100, run:

python main.py --dataset cifar10 (cifar100) --threshold 0.9

The config --threshold 0.9 is used for selecting negative samples to mix.

For graph and node classification. Run:

python main.py

You should download the dataset by yourself.