Efficient Higher Order Self Attention Via Tensor Operations With Topology Enhanced Graph Neural Network For Molecular Graph Classification

Abstract

We modify a recently proposed higher order generalization of self attention mechanism and reduce its time complexity using tensor operations. Then the modified higher order self attention is applied to a graph neural network with a novel structural encoding based on cell complex from algebraic topology. Experiments on molecular graph benchmarks show that our modified higher order self attention is more efficient than the original higher order self attention, and our proposed structural encoding improves the performance of the graph neural network.