Algorithm 1.

ImageGCN batch propagation algorithm.

Require:
	graph node set V and the mini-batch B;
	relation adjacency matrix Ar, r ∈ R;
	input image features X (can be stored externally);
	network depth K;
	number of neighbors to sample for each node and each relation n.
Ensure: The representation of all samples in B
1:	B(K) ← B
2:	for k = K · · · 1 do
3:	B(k−1) ← B(k)
4:	for r ∈ R do
5:	for v ∈ B(k) do
6:	Nr (v) ← the neighbor set of v based on Ar
7:	n ← min(n, |Nr(v)|)
8:	S ← random n samples from Nr(v) without replacement
9:	B(k−1) ← B(k−1) ∪ S
10:	end for
11:	end for
12:	end for
13:	load the features H(0) for B(0) from X
14:	for r ∈ R do
15:	$A_r^B$ ← the sub-matrix corresponding to B(0) in the adjacency matrix Ar
16:	end for
17:	execute Eq. 8 with $A_r=A_r^B$, for k = 0, · · ·, K − 1
18:	extract the representations of samples in B from H(K)