DMH-FSL: Dual-Modal Hypergraph for Few-Shot Learning

Large scale labeled samples are expensive and difficult to obtain, hence few-shot learning (FSL), only needing a small number of labeled samples, is a dedicated technology. Recently, the graph-based FSL approaches have attracted lots of attention. It is helpful to model pair-wise relations among samples according to the similarity of features. However, the data in the reality usually have high-order relations, which can not be modeled by the traditional graph-based methods. To address this challenge, we introduce hypergraph structure and propose the Dual-Modal Hypergraph Few-Shot Learning (DMH-FSL) method to model the relations from different perspectives to model the high-order relations between samples. Specifically, we construct a dual-modal (e.g., feature-modal and label-modal) hypergraph, the feature-modal hypergraph construct incidence matrix with samples' features and the label-modal hypergraph construct incidence matrix with samples' labels. In addition, we employ two hypergraph convolution methods to perform flexible aggregation of samples from different modals. The proposed DMH-FSL method is easy to extend to other graph-based methods. We demonstrate the efficiency of our DMH-FSL method on three benchmark datasets. Our algorithm has at least an increase of 2.62% in Stanford40(from 72.20 to 74.82%), 0.85% in mini-ImageNet(from 50.33 to 51.18%) and 1.61% in USE-PPMI(from 78.77 to 80.38%) in few-shot learning experiments. What's more, the cross-domain experimental results evaluate our method's adaptability in real-world applications to some extent.