We introduce an architecture based on deep hierarchical decompositions to
learn effective representations of large graphs. Our framework extends classic
R-decompositions used in kernel methods, enabling nested "part-of-part"
relations. Unlike recursive neural networks, which unroll a t