Since there lacks aunified standard instoring/saving exemplars forincremental few-shot learning, we choose the setting that we consider most reasonable and practical. In our experiments, we observethat after training on base classes with balanced data, the norms ofthe class prototypes ofbase classes tend tobesimilar. However,after fine-tuning with very few data on unseen new classes, the norms of the new class prototypes are noticeably smaller than those of the base classes. The few-shot novelclasses consist ofhousehold furniture, vehicles2, flowers, and food containers (20 classes in total). The few-shot novel classes consist of people, vehicles2, flowers, and food containers (20 classes in total).

By Assumption 4.1 and 4.2, we have E By Assumption 4.1, an important consequence is that for all {θ,θ Assumption 4.2, we have (see [1, p. 183]) V By Assumption 4.1, 4.2 and 4.3, we have lim inf The first condition in Assumption 4.3 ensures that Under assumptions 4.1, 4.2 and 4.3, we further assume that the risk function Figure 3: Our re-implementation results of Re-balance and ICaRL are very close to those reported in [2]. Mean Standard Deviation Base classes 7.97 0.63 New classes 7.48 0.71 A.3 Additional Experiment Results We re-implement FSLL because the code is not provided. Rebalance are very close to those reported in [2]. To verify the correctness of our implementation of FSLL [3], we compare the results of our implementation and those reported in [3] in Table 8. In our experiments, we observe that after training on base classes with balanced data, the norms of the class prototypes of base classes tend to be similar.

Node classification is a central task in graph data analysis. Scarce or even no labeled data of emerging classes is a big challenge for existing methods. A natural question arises: can we classify the nodes from those classes that have never been seen? In this paper, we study this zero-shot node classification (ZNC) problem which has a two-stage nature: (1) acquiring high-quality class semantic descriptions (CSDs) for knowledge transfer, and (2) designing a well generalized graph-based learning model. For the first stage, we give a novel quantitative CSDs evaluation strategy based on estimating the real class relationships, so as to get the "best" CSDs in a completely automatic way. For the second stage, we propose a novel Decomposed Graph Prototype Network (DGPN) method, following the principles of locality and compositionality for zero-shot model generalization. Finally, we conduct extensive experiments to demonstrate the effectiveness of our solutions.