Saed Rezayi

Thanks for your prompt reply. Actually I'm doing the link prediction task. There is no notion of class in my dataset and all the other settings are similar to yours....

Maybe I'm wrong but in calculating training accuracy you just use the original adjacency graph (`adj_orig`) and its reconstructed version (`model.reconstructions`): [Line 18](https://github.com/tkipf/gae/blob/a1aecb00de894859f509f2d392135086a86946a0/gae/optimizer.py#L18) of optimizer.py: `self.correct_prediction = tf.equal(tf.cast(tf.greater_equal(tf.sigmoid(preds_sub), 0.5), tf.int32),...

Nope, I couldn't figure out why sometimes train_acc is decreasing while val_ap is increasing, and I don't know if this is okay or an issue!!