Graph-based node classification based on connectivity and topology

Abstract

Techniques are disclosed for determining predictions from a graph of a network dataset. The graph of the network dataset may include nodes describing entities and edges describing connections or links between the entities. Predictions may be made using a dual-path convolution network that considers both node connectivity and node topology. Node topology includes assessment of similarities in topology roles between nodes in the graph, even nodes that reside in different parts of the graph. The node connectivity and node topology in the dual-path convolution may be aligned using a multi-head attention network. Outputs from previous layers of the multi-head attention network may be provided as inputs to subsequent layers of the dual-path convolution to mutually reinforce the convolutions determining node connectivity and node topology toward alignment.