This research proposes an Unsupervised Episode Generation method, "Neighbor as Queries (NaQ)", enabling Unsupervised Graph Meta-learning to solve Few-Shot Node Classification (FSNC) tasks.

As it is model-agnostic, any of existing Graph Meta-learning methods based on the episodic learning framework can be trained in unsupervised manner (without labels).

NaQ resolves the following problems of existing approaches for FSNC:

1. Label-scarcity problem of supervised Graph Meta-learning methods.
   Label-scarcity problem hinders the full utilization of information of all nodes in a graph, which leads to overfitting to labeled nodes and causes FSNC performance degradation.

2. Class Imbalance problem of GCL methods with Linear Probing.
   As GCL methods cannot attain the knowledge of what downstream task to solve during the training phase, so its output node embeddings lack generalizability to solve FSNC.
   For this reason, GCL methods exhibit more degraded FSNC performance in more imbalanced settings.

If you are interested in this research, please refer to the following links:
Link: Paper (arXiv) / Code (Github) / Slides / Video (YouTube)