The proposed AvoGCL model generates contrastive views of increasing difficulty through two trainable adversarial perturbators, which significantly outperforms state-of-the-art competitors in recommendation tasks.
