Unverified50% confidenceBenchmarkExact time
On the target logit, 10 of the 11 reportable GCN and GATv2 units show relative fidelity that depends on the evaluation intervention; in 9 of these units, each ranking achieves a larger deletion area under the intervention it was itself constructed for.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
11/4/2026
First Seen
Valid until: 2/2/2027
Sources
Related Entities
Related Claims
UnverifiedGCN's degree-normalization coefficients depend on node degree, and deleting an edge changes the degree and thus the coefficients of the surviving edges, so GCN does not satisfy the equivalence condition of "edge coefficients independent of the edge set."67% similarUnverifiedFor GNN models with additive message passing and edge coefficients independent of the edge set, structural deletion and message suppression are exactly equivalent interventions.63% similarUnverified图神经网络(GNN)核心操作是消息传递,不同图卷积变体如GCN、GAT、GraphSAGE在不同任务上各有优劣60% similarUnverifiedOn the feeder data, parameter sharing recovers about two-thirds of the error gap between node-specific models and the GNN (relative to the regularized baseline).60% similarUnverifiedThe GGUF format stores a single expert count for the entire model, so every layer must retain exactly the same number of experts; imbalance across layers would lead to padding, which would offset the memory savings from pruning.59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1134411API
curl https://kongchang.com/api/v1/knowledge/claims/1134411MCP
get_claim(id=1134411)