Abstract: Highlights•A new model called GIUNet is presented for the representation learning on graphs.•GIUNet exploits structural information and node features during the pooling stage.•GIUNet employs graph isomorphism convolution in encoder–decoder structure.