Algorithm 1: GraphSAGE algorithm. The model first takes the node features as inputs. For each layer, the model aggregates the information from neighbors and then updates the hidden state of each node $v_i$.

Input: $G=(V,E)$: The graph G with set of nodes V and set of edges E.             $x_i$: The input features of node $v_i$             L: The depth of hidden layers, $\forall l\in \{1\cdots L\}$             ${AGG}_k$: Differentiable aggregator functions             $N\left(v_i\right)$: The set of neighbors of node $v_i$. Output: $Z_i$: Vector representations for $v_i$. $h_i^0\leftarrow x_i,\forall v_i\in V$ $Z_i\leftarrow h_i^L,\forall v_i\in V$