On representation power of neural network-based graph embedding and beyond