Appendix A The Proof of Prop. 1 Denote P

Compared with Eq. 25 and Eq. 29, we find We can see the approximate solutions. The tasks are binary classification. We use the default train/val/test split with ratio 8:1:1. The experiments are easily influenced by the parameter initialization. The AUC-ROC results are reported in Tab. 2. H.4 We use the official splits and evaluation is conducted on validation set.