Few-shot learning via dual hypergraph graph convolutional representation

Transductive few-shot learning (FSL), which aims to train a model that can identify the labels of query set samples with limited labeled support set samples is a fundamental problem in computer vision field. The main challenges of FSL are twofold: i) how to effectively utilize the limited labeled support set sample for learning effective data representation for both support and query samples andl ii) how to discover the relationships between the support set and query set samples for query data label prediction. In this paper, we propose to address the above challenge i andl ii by developing Dual Hypergraph Graph Convolutional Representation (DHGCR) models: Support Hypergraph (SHGraph) and Support-Query Hypergraph (SQHGraph) representation modules. To be specific, in SHGraph, nodes denote the samples in support set and the hyper-edges correspond to the different classes. We then adopt a hypergraph convolutional module to learn context-aware representation for each support data sample. After that, we then build the SQHGraph over both query set and support set and perform hypergraph convolution to learn context-aware representations for both support and query samples. Finally, we employ graph-based label propagation for FSL problem. On the mini-ImageNet, tiered-ImageNet, CIFAR-FS, and CUB datasets, we show that the proposed hypergraph representation method outperforms other state-of-the-art methods in transductive FSL tasks.