Geometric algebra graph neural network for cross-domain few-shot classification

Graph neural networks (GNNs) show powerful processing ability on graph structure data for nodes and graph classification. However, existing GNN models may cause information loss with the increasing number of the network layer. To improve the graph-structured data features representation quality, we introduce geometric algebra into graph neural networks. In this paper, we construct a high-dimensional geometric algebra (GA) space in the Non-Euclidean domain to better learn vector embedding for graph nodes. We focus our study on few-shot learning and propose a geometric algebra graph neural network (GA-GNN) as the metric network for cross-domain few-shot classification tasks. In the geometric algebra space, the feature nodes are mapped into hyper-complex vector, which helps reduce the distortion of feature information with the increased hidden layers. The experimental results demonstrate that the approach we proposed achieves the state-of-the-art few-shot cross-domain classification accuracy in five public datasets.