MAEM

MAEM is a replay-free, capacity-controllable dynamic mixture-of-experts framework for class-incremental graph learning.

MAEM uses Simplified Graph Convolution (SGC) to build node representations; SGC is more lightweight and more stable than standard GCN, making it suitable for continual learning scenarios.

In MAEM, whenever a new session arrives, a dedicated expert is trained to handle the current classes, and the expert is frozen immediately after training is complete; new knowledge can only enter new experts, and the parameters of old experts are not shaken.

MAEM equips each expert with a masked autoencoder (MAE), and a test node is assigned to the expert with the lowest reconstruction error; this decentralized routing mechanism requires no task ID and no shared parameters.

When MAEM reaches a fixed capacity limit, it merges the least-used experts via the Git Re-Basin alignment technique, and distills the merged experts into a single unified fused expert using topology-aware pseudo-data synthesized from the MAE distributions.

MAEM was validated on four class-incremental graph learning benchmarks, and the results show that recency bias is significantly alleviated, performance surpasses several strong continual graph learning baselines, and model capacity remains bounded under strictly replay-free conditions.

MAEM does not rely on task identity, does not store historical data, and does not share a trainable router.

At inference time, MAE requires one forward pass per expert, and an increase in the number of experts linearly increases inference overhead.

Git Re-Basin alignment incurs additional computational cost when the number of experts is large, which is an engineering issue that needs to be considered in actual deployment.

The MAEM paper abstract does not disclose the specific capacity threshold and merge trigger frequency, and the sensitivity of these hyperparameters remains to be verified by further experiments.