We present a provably expressive graph learning framework based on representing graphs as bags of subgraphs and processing them with an equivariant architecture, composed of GNNs and set networks.