In supervised learning problems where we need to make graph$-$level predictions $($e$.$g$.$ molecules are represented as graphs, and we construct a graph neural network to predict the melting temperature associated with a given molecule$),$ a pooling function is used to combine information from learned node embeddings that does not depend on the ordering of the nodes $($since properties associated with a graph should not depend on the order in which we number the nodes$).$ Pooling functions such as mean and max satisfy this invariance to node ordering while still allowing the model to distinguish between different graph structures.