label leakage in your code

[Tianbaojie]

hi, your code is causing label leakage. You used the test and validation sets and edge information during training

wandb_train.py self.all_edge_index include test edge,val edge def train(self): time_total_start = time.time() wandb.log({"epochs": self.epoch}) for epoch in range(self.epoch): t = time.time() self.model.train() self.optimizer.zero_grad() adj_recon, mu, logvar = self.model(self.data.x, self.all_edge_index) loss = self.criterion.loss_fn(mu=mu, logvar=logvar, pos_edge_index=self.data.train_pos_edge_index, all_edge_index=self.all_edge_index) loss.backward() self.optimizer.step()

[Anindyadeep]

def train(self):
    time_total_start = time.time()
    wandb.log({"epochs": self.epoch})
    for epoch in range(self.epoch):
        t = time.time()
        self.model.train()
        self.optimizer.zero_grad()
        adj_recon, mu, logvar = self.model(self.data.x, self.all_edge_index)
        loss = self.criterion.loss_fn(mu=mu,
                                logvar=logvar,
                                pos_edge_index=self.data.train_pos_edge_index,
                                all_edge_index=self.all_edge_index)
        loss.backward()
        self.optimizer.step()

        if (epoch + 1) % 10 == 0:
            self.model.eval()
            roc_auc, ap = self.criterion.single_test(
                mu = mu,
                logvar=logvar,
                test_pos_edge_index=self.data.test_pos_edge_index,
                test_neg_edge_index=self.data.test_neg_edge_index
            )
            self.__metric_log(
                loss = loss.item(),
                roc = roc_auc,
                ap = ap,
                time_lapsed = time.time() - t
            )
            epoch_str = str(epoch + 1).zfill(3) if epoch + 1 < 100 else str(epoch + 1)
            time_elapsed = time.time() - t

            formatted_output = (
                f"TESTING | AT epoch {epoch_str}, "
                f"loss: {loss.item():.4f}, "
                f"ROC_AUC: {roc_auc:.4f}, "
                f"AP: {ap:.4f}, "
                f"Time: {time_elapsed:.4f}"
            )

            print(formatted_output)

    print(f"The total time taken: {(time.time() - time_total_start):.4f}")

Hey, sorry for the very late reply. But, if all the train and test samples are inside the data object. Now here, I might provide all the edge indices but also at the same time, I am providing the train position indices that does the required masking to not leak any evaluation data. However this is also a problem in transductive learning, where we can not do complete separation of training data while we are doing evaluation as there is a dependency of that.

[Anindyadeep]

Making this issue close due to no activity