Graph Pointer Neural Networks

Graph Neural Networks (GNNs) have shown advantages in various graph-based applications. Most existing GNNs assume strong homophily of graph structure and apply permutation-invariant local aggregation neighbors to learn a representation for each node. However, they fail generalize heterophilic graphs, where most neighboring nodes different labels or features, the relevant are distant. Few recent studies attempt address this problem by combining multiple hops hidden representations central (i.e., multi-hop-based approaches) sorting based on attention scores ranking-based approaches). As result, these approaches some apparent limitations. On one hand, do not explicitly distinguish from large number multi-hop neighborhoods, leading severe over-smoothing problem. other models joint-optimize node ranking with end tasks result sub-optimal solutions. In work, we present Pointer (GPNN) tackle challenges mentioned above. We leverage pointer network select amount which constructs an ordered sequence according relationship 1D convolution is then applied extract high-level features sequence. The pointer-network-based ranker GPNN joint-optimized parts end-to-end manner. Extensive experiments conducted six public classification datasets graphs. results show that significantly improves performance state-of-the-art methods. addition, analyses also reveal privilege proposed filtering out irrelevant reducing over-smoothing.