MAGIC: Topology-Aware Analytic Graph Few-Shot Class-Incremental Learning

Junlin Chen; Yuhan Wang; Xuefei Wang; Xiao Wang; Ruijie Wang; Jianxin Li

cs.LGOct 4, 2026

MAGIC: Topology-Aware Analytic Graph Few-Shot Class-Incremental Learning

Authors: Junlin Chen, Yuhan Wang, Xuefei Wang, Xiao Wang, Ruijie Wang, Jianxin Li

Organizations: Beihang University, Beijing, China

Abstract

Graph few-shot class-incremental learning (GFSCIL) requires a model to continually recognize emerging classes from only a few labeled nodes while preserving previously acquired knowledge. Beyond the catastrophic forgetting inherited from conventional graph continual learning, GFSCIL presents two distinctive challenges: extremely limited novel-class supervision causes severe overfitting, while cross-session edges---edges connecting newly arriving nodes with historical nodes---alter historical propagation neighborhoods and thereby induce representation drift. We propose MAGIC, a replay-free GFSCIL framework that combines a frozen graph representation backbone (e.g., an intrinsically parameter-free backbone such as SGC or a pretrained graph foundation model) with closed-form analytic continual learning. To alleviate novel-session overfitting, MAGIC learns a topological prior from the base graph that can characterize both homophilous and heterophilous relations, and injects this prior through Potts Markov random field inference to refine supervision for novel classes. To mitigate representation drift, MAGIC transfers previous predictions from the old representations of affected historical nodes to their updated representations through drift-aware analytic distillation. Experiments across five datasets and eight baselines demonstrate the effectiveness of MAGIC. Under the 5-shot setting, MAGIC improves Mean Accuracy and Final Accuracy by 5.48 percentage points and 9.33 percentage points on average, and reduces Performance Drop by 10.78 percentage points on average compared with the best baselines. MAGIC also shows clear advantages under the 1- and 3-shot settings, with larger gains as the number of supports increases. Moreover, MAGIC requires substantially less training time.

Figures & tables

Figure 1: On the base stage, MAGIC learns a topological prior. In novel sessions, limited labeled supports first update an initial classifier, while drift-aware distillation protects historical nodes affected by cross-session edges. Finally, MAGIC extracts a subgraph around novel supports and performs Potts MRF inference with the topological prior to adjust the pseudo-labels predicted by the initial classifier, and subsequently updates the classifier using these adjusted labels.

	CoraFull			ogbn-arxiv			Coauthor CS			Amazon Computers			WikiCS
Method	MA	FA	PD	MA	FA	PD	MA	FA	PD	MA	FA	PD	MA	FA	PD
LwF	16.06	3.82	79.86	18.16	4.65	82.03	35.83	17.16	79.73	35.05	13.55	82.07	35.83	12.90	79.78
LwF (std)	$\pm$ 1.24	$\pm$ 1.44	$\pm$ 2.35	$\pm$ 0.97	$\pm$ 5.80	$\pm$ 6.57	$\pm$ 3.44	$\pm$ 5.45	$\pm$ 5.73	$\pm$ 1.73	$\pm$ 5.61	$\pm$ 4.53	$\pm$ 2.33	$\pm$ 3.96	$\pm$ 4.33
EWC	16.14	3.86	79.82	15.87	6.07	80.52	36.38	17.32	79.58	35.13	13.96	81.54	35.89	12.86	79.82
EWC (std)	$\pm$ 1.31	$\pm$ 1.39	$\pm$ 2.30	$\pm$ 2.23	$\pm$ 6.70	$\pm$ 7.19	$\pm$ 3.85	$\pm$ 5.54	$\pm$ 5.78	$\pm$ 1.74	$\pm$ 5.92	$\pm$ 4.90	$\pm$ 2.49	$\pm$ 3.86	$\pm$ 4.27
MAS	16.31	6.07	77.61	34.40	27.37	59.25	45.91	20.77	76.12	34.98	13.35	82.17	35.24	12.42	80.26

Table 1: Main results under the 5-shot setting. MAGIC Backbone: SGC. MAGIC improves Mean Accuracy and Final Accuracy by 5.48 percentage points and 9.33 percentage points on average, and reduces Performance Drop by 10.78 percentage points on average compared with the best baselines.

Figure 4: Main results on CoraFull under the 1- and 3-shot settings. MAGIC Backbone: SGC

	CoraFull			ogbn-arxiv			Coauthor CS			Amazon Computers			WikiCS
Backbone	MA	FA	PD	MA	FA	PD	MA	FA	PD	MA	FA	PD	MA	FA	PD
SGC	65.45	56.02	24.19	63.50	49.81	34.37	94.03	91.45	5.54	86.54	81.04	9.70	80.33	74.64	14.09
SGC (std)	$\pm$ 0.95	$\pm$ 1.05	$\pm$ 3.65	$\pm$ 3.92	$\pm$ 2.35	$\pm$ 2.55	$\pm$ 0.79	$\pm$ 0.28	$\pm$ 1.34	$\pm$ 1.68	$\pm$ 2.88	$\pm$ 5.06	$\pm$ 1.74	$\pm$ 2.14	$\pm$ 3.01
S 2 GC	64.42	54.74	25.48	61.23	47.05	36.43	94.25	91.77	5.43	87.26	81.69	9.76	80.61	74.86	14.20
S 2 GC (std)	$\pm$ 0.57	$\pm$ 1.20	$\pm$ 4.24	$\pm$ 4.01	$\pm$ 2.39	$\pm$ 2.66	$\pm$ 0.78	$\pm$ 0.48	$\pm$ 1.40	$\pm$ 1.58	$\pm$ 2.72	$\pm$ 4.79	$\pm$ 1.55	$\pm$ 2.09	$\pm$ 3.34
PPR	66.92	57.62	22.84	65.40	51.34	34.13	94.26	91.90	5.02	85.89	79.96	10.58	80.12	73.90	14.98

Table 2: Results (%) with different frozen graph backbones under the 5-shot setting. RGVT is excluded from ogbn-arxiv because its released checkpoint was supervised-pretrained on that dataset.

Variant	Mean Acc.	Final Acc.	PD
w/o Prior	53.69 $\pm$ 7.74	37.29 $\pm$ 7.28	46.89 $\pm$ 6.72
w/o Distill	61.88 $\pm$ 3.61	48.02 $\pm$ 1.69	36.16 $\pm$ 4.61
MAGIC	63.50 $\pm$ 3.92	49.81 $\pm$ 2.35	34.37 $\pm$ 2.55

Table 3: Ablation results (%) on ogbn-arxiv under the 5-shot setting using SGC.

Figure 5: Novel-class performance on CoraFull.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Nodes	Edges	Features	Classes	$h_{\mathrm{edge}}$
CoraFull	19,793	126,842	8,710	70	0.57
Coauthor CS	18,333	163,788	6,805	15	0.81
ogbn-arxiv	169,343	1,166,243	128	40	0.65
WikiCS	11,701	216,123	300	10	0.65
Amazon Computers	13,752	491,722	767	10	0.78

Appendix

Table 4: Statistics of the node-classification datasets.

Dataset	Base Classes	Way
CoraFull	20	5
ogbn-arxiv	10	3
Coauthor CS	5	2
Amazon Computers	4	2
WikiCS	4	2

Appendix

Table 5: Class-incremental split configurations.

	CoraFull			ogbn-arxiv			Coauthor CS			Amazon Computers			WikiCS
Method	MA	FA	PD	MA	FA	PD	MA	FA	PD	MA	FA	PD	MA	FA	PD
LwF	13.18	2.28	81.39	16.64	5.29	81.38	35.23	16.68	80.21	32.84	13.17	82.40	34.56	11.45	81.14
LwF (std)	$\pm$ 0.93	$\pm$ 0.93	$\pm$ 2.91	$\pm$ 1.53	$\pm$ 6.77	$\pm$ 7.35	$\pm$ 3.34	$\pm$ 5.40	$\pm$ 5.82	$\pm$ 2.10	$\pm$ 5.34	$\pm$ 4.25	$\pm$ 2.40	$\pm$ 4.66	$\pm$ 4.76
EWC	13.53	2.83	80.85	17.48	9.11	77.57	36.50	17.14	79.76	33.16	14.63	80.92	34.83	11.76	80.91
EWC (std)	$\pm$ 1.09	$\pm$ 1.71	$\pm$ 2.75	$\pm$ 1.25	$\pm$ 5.30	$\pm$ 5.67	$\pm$ 4.02	$\pm$ 5.25	$\pm$ 5.64	$\pm$ 2.62	$\pm$ 6.77	$\pm$ 5.92	$\pm$ 2.48	$\pm$ 4.98	$\pm$ 5.12
MAS	16.40	6.16	77.52	30.77	18.51	68.16	51.31	24.55	72.34	34.88	12.71	82.95	35.65	12.35	80.26

Appendix

Table 6: Main results (%) under the 1-shot setting. The mean and standard deviation over five runs are reported in consecutive rows. The best result in each column is shown in bold. MAGIC Backbone: SGC.

	CoraFull			ogbn-arxiv			Coauthor CS			Amazon Computers			WikiCS
Method	MA	FA	PD	MA	FA	PD	MA	FA	PD	MA	FA	PD	MA	FA	PD
LwF	15.46	3.51	80.17	18.14	5.29	81.33	35.72	16.99	79.90	34.54	13.39	82.23	35.02	12.51	80.11
LwF (std)	$\pm$ 1.24	$\pm$ 1.15	$\pm$ 2.46	$\pm$ 0.66	$\pm$ 6.95	$\pm$ 7.53	$\pm$ 3.54	$\pm$ 5.19	$\pm$ 5.42	$\pm$ 2.29	$\pm$ 5.47	$\pm$ 4.37	$\pm$ 0.70	$\pm$ 3.58	$\pm$ 3.92
EWC	15.56	3.61	80.07	16.05	6.13	80.49	36.66	17.07	79.82	35.00	14.46	81.12	35.16	12.57	80.11
EWC (std)	$\pm$ 1.36	$\pm$ 1.45	$\pm$ 2.43	$\pm$ 2.05	$\pm$ 6.67	$\pm$ 7.13	$\pm$ 4.20	$\pm$ 5.07	$\pm$ 5.32	$\pm$ 2.59	$\pm$ 6.46	$\pm$ 5.58	$\pm$ 0.80	$\pm$ 3.74	$\pm$ 4.14
MAS	16.04	6.08	77.60	32.89	27.88	58.80	49.00	21.31	75.58	34.83	12.54	83.04	35.23	12.35	80.33

Appendix

Table 7: Main results (%) under the 3-shot setting. The mean and standard deviation over five runs are reported in consecutive rows. The best result in each column is shown in bold. MAGIC Backbone: SGC.

	CoraFull		ogbn-arxiv		Coauthor CS		Amazon Computers		WikiCS
Method	AN	FN	AN	FN	AN	FN	AN	FN	AN	FN
LwF	16.22	3.27	11.26	6.84	43.70	22.85	47.56	33.32	46.27	23.89
LwF (std)	$\pm$ 2.05	$\pm$ 1.27	$\pm$ 4.13	$\pm$ 8.03	$\pm$ 2.48	$\pm$ 9.78	$\pm$ 10.84	$\pm$ 13.99	$\pm$ 4.14	$\pm$ 9.25
EWC	16.93	3.94	13.75	10.75	43.77	23.17	46.60	33.48	47.07	24.47
EWC (std)	$\pm$ 1.67	$\pm$ 2.26	$\pm$ 3.60	$\pm$ 6.83	$\pm$ 3.45	$\pm$ 9.73	$\pm$ 11.08	$\pm$ 13.98	$\pm$ 3.77	$\pm$ 9.73
MAS	20.32	8.79	7.99	8.17	58.28	30.91	59.58	32.25	50.77	26.07

Appendix

Table 8: Novel-class results (%) under the 1-shot setting. The mean and standard deviation over five runs are reported in consecutive rows. The best result in each column is shown in bold. MAGIC Backbone: SGC.

	CoraFull		ogbn-arxiv		Coauthor CS		Amazon Computers		WikiCS
Method	AN	FN	AN	FN	AN	FN	AN	FN	AN	FN
LwF	22.19	4.99	15.97	6.85	45.59	23.23	57.41	33.81	49.51	26.29
LwF (std)	$\pm$ 1.20	$\pm$ 1.35	$\pm$ 3.74	$\pm$ 8.27	$\pm$ 3.44	$\pm$ 9.26	$\pm$ 12.08	$\pm$ 14.12	$\pm$ 9.38	$\pm$ 6.89
EWC	22.33	5.10	17.01	7.89	45.86	23.22	58.02	33.94	50.15	26.40
EWC (std)	$\pm$ 1.46	$\pm$ 1.80	$\pm$ 3.79	$\pm$ 7.96	$\pm$ 3.43	$\pm$ 9.13	$\pm$ 12.46	$\pm$ 14.22	$\pm$ 9.32	$\pm$ 7.20
MAS	22.38	8.59	9.94	8.05	57.42	28.54	59.77	31.85	49.37	26.04

Appendix

Table 9: Novel-class results (%) under the 3-shot setting. The mean and standard deviation over five runs are reported in consecutive rows. The best result in each column is shown in bold. MAGIC Backbone: SGC.

	CoraFull		ogbn-arxiv		Coauthor CS		Amazon Computers		WikiCS
Method	AN	FN	AN	FN	AN	FN	AN	FN	AN	FN
LwF	23.86	5.41	15.06	6.01	46.25	23.51	60.18	34.25	51.46	27.22
LwF (std)	$\pm$ 0.88	$\pm$ 1.72	$\pm$ 3.72	$\pm$ 6.89	$\pm$ 4.24	$\pm$ 9.79	$\pm$ 8.12	$\pm$ 14.62	$\pm$ 7.31	$\pm$ 8.14
EWC	23.88	5.47	16.26	7.57	46.48	23.71	59.76	34.16	51.65	27.16
EWC (std)	$\pm$ 0.95	$\pm$ 1.65	$\pm$ 3.72	$\pm$ 8.06	$\pm$ 4.32	$\pm$ 9.88	$\pm$ 7.98	$\pm$ 14.46	$\pm$ 7.30	$\pm$ 8.00
MAS	23.51	8.60	10.06	7.04	55.78	28.19	60.05	33.65	49.31	26.31

Appendix

Table 10: Novel-class results (%) under the 5-shot setting. The mean and standard deviation over five runs are reported in consecutive rows. The best result in each column is shown in bold. MAGIC Backbone: SGC.

Figure 6: Hyperparameter sensitivity of MAGIC on CoraFull under the 5-shot setting. Each subplot varies one hyperparameter while fixing all others to their default values. MAGIC Backbone: SGC

Dataset	$\eta$	$\rho$	$\kappa$	$B$	$T_{D}$	$\tau_{u}$	$k$
CoraFull	3.0	1.0	0.2	3	0.50	1.0	3
ogbn-arxiv	6.0	1.0	0.5	10	0.25	1.0	3
Coauthor CS	1.0	1.0	0.5	10	0.25	1.0	3
Amazon Computers	1.0	1.0	0.5	10	0.25	1.0	3
WikiCS	3.0	1.0	0.5	10	0.25	1.0	3

Appendix

Table 11: Hyperparameter configurations used in the main experiments. The same configurations are used for the 1-, 3-, and 5-shot settings.

Explore similar work

May 23, 2026

Advancing Graph Few-Shot Learning via In-Context Learning

Graph few-shot learning, which aims to classify nodes from novel classes with only a few labeled examples, is a widely studied problem in graph learning. However, existing methods often face two key limitations. First, the predominant graph few-shot learning paradigm relies on supervised tasks, failing to leverage the vast number of unlabeled nodes in the graph. Second, many approaches require complex task adaptation or fine-tuning during inference, limiting their efficiency and applicability. Inspired by the powerful in-context learning capabilities of large language models, we propose a novel model named VISION for adVancIng graph few-Shot learning via In-cOntext LearNing to address these challenges. Our model reframes graph few-shot learning as a fine-tuning-free sequence reasoning problem. At its core is a context-aware network that initializes nodes with role embeddings and employs a dual-context fusion module to synergistically integrate local topological structures and global task-level dependencies. This allows our model to dynamically generate class-aware representations for the query set conditioned on the support set context in a single forward pass. To effectively train our model, we introduce an unsupervised task generator that creates structure-adaptive features and constructs diverse pseudo-tasks from abundant unlabeled data. Our method unifies unsupervised meta-learning with graph in-context learning, achieving efficient inference. Extensive experiments on multiple benchmark datasets demonstrate the superiority of our model. Our public code can be found

Renchu Guan, Yajun Wang, Chunli Guo +5

Key Laboratory of Symbolic Computation and Knowledge Engineering of the Ministry of Education

Meta-Learning Few-Shot Learning In-Context Learning

Ask PDF

Oct 14, 2026

MAEM: Mitigating Recency Bias via a Replay-Free, Capacity-Controlled Dynamic Mixture of Experts for Class-Incremental Graph Learning

Graph Neural Networks (GNNs) excel on graph-structured data but suffer catastrophic forgetting when classes arrive sequentially. In class-incremental graph learning, predictions are often dominated by recently learned classes, causing severe recency bias. Dynamic expansion can isolate knowledge across sessions, but many methods assume task-incremental settings and require task identities at test time to select branches or heads---information unavailable in class-incremental learning. Learnable routers offer an alternative, yet continual fine-tuning the router on new data can bias routing toward recent components. We propose MAEM, a replay-free, capacity-controlled dynamic mixture-of-experts framework for class-incremental graph learning. MAEM builds representations with Simplified Graph Convolution (SGC) and trains a dedicated expert for each new session, then freezes it to protect prior knowledge. For routing without task IDs or a shared trainable router, each expert uses a masked autoencoder (MAE); test nodes are assigned to the expert with the lowest reconstruction error, reducing routing-level recency bias. When the expert pool reaches a fixed capacity, MAEM merges the least-used experts via Git Re-Basin alignment and further distills them into a fused expert using topology-aware pseudo data synthesized from MAE distributions (DeepInversion-style), without storing or revisiting historical graph data. On four class-incremental graph benchmarks, MAEM mitigates recency bias, outperforms strong continual graph learning baselines, and keeps model capacity bounded under strict replay-free learning.

Junlin Chen, Wenhao Wang, Junchao Xu +5

Beihang University, Beijing, China · Renmin University of China, Beijing, China · Trine University, Indiana, USA +1

Class-Incremental Learning Graph Neural Networks Mixture of Experts

Ask PDF

Oct 4, 2026

Learning to Curate What You Generate for Generalizable Few-Shot Class-Incremental Learning

Few-shot class-incremental learning (FSCIL) aims to learn novel classes from limited annotations while preserving prior knowledge. Existing methods typically assume a sufficiently large base session, but this assumption fails when both base and incremental data are scarce, leading to weak initial representations, semantic drift, and unstable boundaries. We study this underexplored yet realistic setting, termed Generalizable FSCIL (G-FSCIL), where the base session itself contains only a few classes. Although synthetic data can alleviate supervision scarcity, naively mixing generated samples often introduces semantic noise and exacerbates old-new boundary conflicts. To address this, we propose a framework that curates trustworthy synthetic knowledge for stable G-FSCIL. Specifically, we first construct class-specific synthetic candidate pools using a frozen latent diffusion model, where class inversion is performed at the first observation and the resulting condition embeddings are reused for on-demand generation. Building on these candidates, we learn a knowledge curation strategy that selects samples with both semantic consistency and visual diversity, and distill this process into a transferable selection policy during the base session, which is then reused without further optimization. Leveraging the curated synthetic data, we further design a boundary-stable incremental adaptation scheme, including synthetic-informed prototype initialization and bidirectional boundary calibration to mitigate old-new conflicts. Extensive experiments demonstrate that our method consistently outperforms existing FSCIL baselines, with reduced forgetting and improved balance between old and new classes. Code is available at https://github.com/NiHaoWoJiaoYYC/G-FSCIL.

Junhui Yin, Yuchen Yang, Yilin Yin +6

Beijing University of Posts and Telecommunications, Beijing, China · Institute of Automation, Chinese Academy of Sciences, Beijing, China · Singapore Management University, Singapore, Singapore

Class-Incremental Learning Data Selection Synthetic Data Augmentation

Ask PDF