핵심 개념
The proposed global-local anchor representation (GLAR) learning method can efficiently perform inductive reasoning on opening subgraphs and learn rich entity-independent features for emerging entities in knowledge graphs.
초록
The paper proposes a novel global-local anchor representation (GLAR) learning method for inductive knowledge graph completion (KGC). Unlike previous methods that utilize enclosing subgraphs, GLAR extracts a shared opening subgraph for all candidate entities and performs reasoning on it, enabling the model to reason more efficiently.
The key components of GLAR are:
-
Local Anchor Representation Learning:
- GLAR extracts an opening subgraph around the query entity and defines local anchors as the center node and its one-hop neighbors.
- The nodes in the subgraph are labeled based on these local anchors to capture rich local structure features.
-
Global Anchor Representation Learning:
- GLAR selects global anchors by clustering the nodes based on their neighboring relation features.
- The nodes are further labeled using these global anchors to learn entity-independent global structure features.
-
Global-Local Graph Reasoning:
- A global-local graph neural network is applied to collaboratively propagate both local and global neighborhood features for effective inductive reasoning.
The experiments on three benchmark inductive KGC datasets demonstrate that GLAR outperforms state-of-the-art methods in terms of both ranking and classification metrics.
통계
The number of relations in the training and test KGs ranges from 9 to 222.
The number of entities in the training and test KGs ranges from 922 to 7208.
The number of triples in the training and test KGs ranges from 1034 to 33916.
인용구
"Unlike previous methods that utilize enclosing subgraphs, we extract a shared opening subgraph for all candidates and perform reasoning on it, enabling the model to perform reasoning more efficiently."
"We design some transferable global and local anchors to learn rich entity-independent features for emerging entities."
"Extensive experiments show that our GLAR outperforms most existing state-of-the-art methods."