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

Junlin Chen; Wenhao Wang; Junchao Xu; Qingsong Yao; Fei Xue; Li Sun; Xiao Wang; Ruijie Wang

cs.LGOct 14, 2026

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

Authors: Junlin Chen, Wenhao Wang, Junchao Xu, Qingsong Yao, Fei Xue, Li Sun, Xiao Wang, Ruijie Wang

Organizations: Beihang University, Beijing, China · Renmin University of China, Beijing, China · Trine University, Indiana, USA · Beijing University of Posts and Telecommunications, Beijing, China

Abstract

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.

Figures & tables

Figure 1: Session-wise accuracy matrix of LwF on cora-full. The entry in row $i$ and column $j$ denotes the accuracy on session $j$ after training through session $i$ .

Figure 2: Overall framework of MAEM.

		OGBN-Arxiv			Cora-Full
Method	Replay	Micro	AP	AF	Micro	AP	AF
Bare	$\times$	$0.52\pm 0.24$	$6.99\pm 0.46$	$-83.71\pm 0.87$	$1.60\pm 0.16$	$6.86\pm 0.48$	$-90.99\pm 0.69$
EWC	$\times$	$0.39\pm 0.01$	$6.90\pm 0.17$	$-75.86\pm 1.59$	$2.02\pm 0.31$	$7.38\pm 0.62$	$-90.59\pm 0.64$
MAS	$\times$	$0.37\pm 0.01$	$6.61\pm 0.20$	$-75.58\pm 1.16$	$1.18\pm 0.03$	$5.53\pm 0.13$	$-91.26\pm 0.30$
LwF	$\times$	$0.72\pm 0.22$	$6.78\pm 0.35$	$-81.97\pm 1.84$	$2.03\pm 0.17$	$7.99\pm 0.55$	$-87.92\pm 0.96$
TWP	$\times$	$0.37\pm 0.02$	$6.54\pm 0.30$	$-76.68\pm 1.41$	$1.10\pm 0.03$	$4.89\pm 0.11$	$-87.44\pm 0.56$

Table 1: Main results on four class-incremental graph benchmarks. Results are percentages reported as mean

\pm

standard deviation over five runs.

\checkmark

and

\times

indicate whether historical information is replayed. MAEM N and MAEM M denote MAEM without and with expert merging, respectively. Bold and underlined values indicate the best and second-best results.

Figure 3: Final Micro accuracy, AP, and routing accuracy under different maximum numbers of experts on OGBN-Arxiv and Cora-Full.

Figure 4: Session-wise Micro accuracy and AP on Cora-Full under different expert-capacity limits.

Variant	Micro	AP
Full model	$\mathbf{67.40\pm 0.28}$	$\mathbf{63.04\pm 0.25}$
Random initialization	$61.13\pm 0.30$	$45.87\pm 0.21$
w/o distillation	$60.21\pm 0.39$	$45.20\pm 0.19$
w/o topology-aware loss	$60.77\pm 0.30$	$45.71\pm 0.21$

Table 2: Ablation results (%) on Cora-Full.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Nodes	Edges	Feature Dimension	Classes	Class per Session	Sessions
ogbn-arxiv	169,343	1,166,243	128	40	4	10
Cora-Full	19,793	65,311	512 †	70	4	17
Coauthor-CS	18,333	81,894	256 †	15	2	7
WikiCS	11,701	216,123	300	10	2	5

Appendix

Table 3: Dataset statistics. “Feat.” is the feature dimension after preprocessing; † denotes Truncated-SVD reduction (original dimensions: 8,710 for Cora-Full and 6,805 for Coauthor-CS ). “

C

/Sess.” is the number of classes per session. For Coauthor-CS , the final incomplete session is discarded, so 14 of 15 classes are used.

Figure 5: Hyperparameter sensitivity on Cora-Full. Left: MAE mask ratio. Right: number of pseudo samples per class used in expert fusion. Curves report Micro and AP (mean

\pm

std over five seeds).

	OGBN-Arxiv			Cora-Full
Backbone	Micro	AP	AF	Micro	AP	AF
Identity	$33.18\pm 1.65$	$35.12\pm 0.20$	$-19.06\pm 1.39$	$43.26\pm 0.25$	$26.89\pm 0.30$	$-1.63\pm 0.27$
SGC / GCN	$\mathbf{50.34\pm 1.41}$	$46.35\pm 0.27$	$\mathbf{-15.04\pm 0.68}$	$68.22\pm 0.08$	$65.54\pm 0.25$	$-7.85\pm 0.36$
GraphSAGE	$43.24\pm 2.03$	$45.64\pm 0.36$	$-18.43\pm 1.14$	$\mathbf{69.72\pm 0.11}$	$65.25\pm 0.25$	$-7.04\pm 0.34$
Random Walk	$48.45\pm 1.60$	$\mathbf{47.04\pm 0.43}$	$-15.75\pm 1.15$	$67.42\pm 0.20$	$\mathbf{65.86\pm 0.17}$	$-9.33\pm 0.44$
PPR	$\underline{49.52\pm 1.56}$	$\underline{46.77\pm 0.24}$	$\underline{-15.63\pm 0.50}$	$\underline{69.69\pm 0.17}$	$65.16\pm 0.16$	$\mathbf{-6.02\pm 0.21}$

Appendix

Table 4: Backbone ablation of MAEM (%). All operators are parameter-free; only the propagation rule changes. Mean

\pm

std over five seeds. Best results are in bold and second-best are underlined.

Expert	Classes	$\mathrm{MMD}^{2}$	$\mathrm{Acc}_{\mathrm{pseudo}}$	$\mathrm{Acc}_{\mathrm{real}}$	Ratio
0	$0$ – $3$	$0.239\pm 0.019$	$0.734\pm 0.087$	$0.903\pm 0.001$	$0.813\pm 0.097$
1	$4$ – $7$	$0.170\pm 0.034$	$0.835\pm 0.017$	$0.909\pm 0.002$	$0.918\pm 0.018$
2	$8$ – $11$	$0.147\pm 0.029$	$0.850\pm 0.027$	$0.955\pm 0.002$	$0.891\pm 0.029$
3	$12$ – $15$	$0.187\pm 0.032$	$0.837\pm 0.017$	$0.941\pm 0.003$	$0.890\pm 0.019$
4	$16$ – $19$	$0.262\pm 0.060$	$0.858\pm 0.077$	$0.932\pm 0.008$	$0.921\pm 0.083$
5	$20$ – $23$	$0.162\pm 0.023$	$0.803\pm 0.022$	$0.904\pm 0.011$	$0.888\pm 0.022$

Appendix

Table 5: Per-expert pseudo-embedding quality on ogbn-arxiv (no-fusion,

K{=}10

\mathrm{MMD}^{2}

uses a Gaussian kernel with median-heuristic bandwidth and an unbiased estimator.

\mathrm{Ratio}=\mathrm{Acc}_{\mathrm{pseudo}}/\mathrm{Acc}_{\mathrm{real}}

under an identical classifier. Mean

\pm

std over five seeds.

Explore similar work

MAGIC: Topology-Aware Analytic Graph Few-Shot Class-Incremental Learning
Oct 4, 2026Junlin Chen, Yuhan Wang, Xuefei Wang +3Class-Incremental Learning Continual Learning
Beihang University, Beijing, China
A Lightweight Plastic-Memory Framework for Graph Few-Shot Class-Incremental Learning
Sep 22, 2026Zihan Mei, Zhili Qin, Tongze Zhang +3Class-Incremental Learning Few-Shot Learning
University of Electronic Science and Technology of China
Analytic Drift Resister for Non-Exemplar Continual Graph Learning
Apr 3, 2026Lei Song, Shihan Guan, Youyong KongIncremental Learning Class-Incremental Learning
School of Computer Science and Engineering, Southeast University, Nanjnig 211189, China

Oct 4, 2026J/K move · Enter open · S save

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

Junlin Chen, Yuhan Wang, Xuefei Wang, Xiao Wang+2

Beihang University, Beijing, China

Class-Incremental Learning Continual Learning Incremental Learning Few-Shot Learning

Open paper AskPDF arXiv

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.

Abstract

Explore similar work

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

A Lightweight Plastic-Memory Framework for Graph Few-Shot Class-Incremental Learning

Analytic Drift Resister for Non-Exemplar Continual Graph Learning