A replay-free method for continual graph learning kept no past graph data, matched the strongest replay baseline that was allowed task identifiers, and ran 5.9 to 14 times faster
AspireAI community · arXiv (cs.LG, cs.DB)
Continual graph learning is supposed to let a model track a graph as it changes without retraining from scratch, but most methods either replay stored subgraphs or need to be told which task a test graph belongs to. CoPE-CGL takes a different route: a parameter-free structural prior, precomputed by multi-hop Laplacian smoothing, anchors learning, memory and inference, so each task adds only a constant amount of storage and no past graph data is retained.
The mechanism is topology-infused bottleneck adaptation plus community-level prototypes, with routing done by majority vote over node-to-prototype matches. At equal storage and without task identifiers, the method reports average performance of 95.17 and 95.72 percent on Arxiv and CoraFull against 94.28 and 95.82 percent for the strongest replay baseline even when that baseline receives true task identifiers, with 5.9 to 14.0 times faster runtime and zero forgetting. Across four transductive and three inductive benchmarks plus mixed streams, it exceeds the second-best of up to eleven baselines by up to 1.41, 6.32 and 9.33 AP percentage points, runs 1.3 to 7.9 times faster end-to-end than the most accurate competitor on each benchmark, and its node-level variant posts the best AP on graphs mixing two tasks.
Two constraints are worth carrying. The authors say plainly that the design targets homophilic graphs with community structure, so the routing-by-prototype mechanism is not expected to transfer unchanged to graphs without that shape. And the comparison set is whatever was publicly available - up to eleven baselines - measured in the authors' own runs, with the default protocol assuming each test graph comes from a single learned task and a node-level variant handling mixed-task graphs.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.