API-GNN: attribute preserving oriented interactive graph neural network

Abstract Attributed graph embedding aims to learn node representation based on the topology and attributes. The current mainstream GNN-based methods of target by aggregating attributes its neighbor nodes. These still face two challenges: (1) In neighborhood aggregation procedure, each would be propagated neighborhoods which may cause disturbance original over-smoothing in GNN iteration. (2) Because is derived from neighbors, topological information have different effects node. However, this contribution has not been considered existing methods. paper, we propose a novel model named API-GNN ( A ttribute P reserving Oriented I nteractive G raph N eural etwork). can only reduce attribute node, but also explicitly impacts representation. We conduct experiments six public real-world datasets validate classification link prediction. Experimental results show that our outperforms several strong baselines over various multiple analysis tasks.