LGLNN: Label Guided Graph Learning-Neural Network for few-shot learning

Abstract

Graph Neural Networks (GNNs) have been employed for few-shot learning (FSL) tasks. The aim of GNN based FSL is to transform the few-shot learning problem into a graph node classification or edge labeling tasks, which can thus fully explore the relationships among samples in support and query sets. However, existing works generally consider the graph learned by node features which ignore the initial pairwise label constraints and thus are generally not guaranteed to be optimal for FSL tasks. Also, existing works generally learn graph edges independently based on node's own features which lack of considering the consistent relationships among different edges. To address these issues, we propose a novel Label Guided Graph Learning-Neural network (LGLNN) model for FSL tasks. The aim of LGLNN is to incorporate the label information to learn an optimal metric graph for GNN by employing the pairwise constraint propagation. The main advantage of LGLNN is that it can learn the metrics (both similarity and dissimilarity) for each graph edge by aggregating the metric information from its neighboring edges and thus can conduct metric learning of all edges cooperatively and consistently. Experimental results demonstrate the effectiveness and better performance of the proposed LGLNN method.

Keywords: Few-shot learning; Graph learning; Graph neural network; Pairwise constraint propagation.