A Hierarchical Mixture-Of-Experts Framework for Few Labeled Node Classification

Few Labeled Node Classification (FLNC) is a challenging subtask of node classification, where training nodes are extremely limited, often with only one or two labels per class. While Graph Neural Networks (GNNs) show promise, they often suffer from feature convergence. A common method to address this challenge is multi-perspective feature extraction, with the Mixture of Experts (MoE) model being a popular approach. However, directly applying MoE to FLNC frequently results in overfitting. To address these issues, we propose the Hierarchical Mixture-of-Experts (HMoE) framework. First, we mitigate overfitting by applying three data augmentation techniques to enrich input features. Next, we design a novel hierarchical mixture-of-experts encoder to achieve diversified feature representations, where the first layer extracts unique feature information, and the second layer refines shared information. Additionally, we design an auxiliary task to distinguish between original and augmented data, using a gradient reversal mechanism to enhance the feature representation ability of graph data. The experimental results show that HMoE outperforms the baseline methods, achieving an average 1.2% performance improvement across six datasets.