ARLIE: Adaptive Reinforcement Learning With Inductive Embeddings for Fully-Inductive Multi-Hop Reasoning Over Temporal Knowledge Graphs

TKG incompleteness: Temporal knowledge graph reasoning addresses the intrinsic incompleteness of temporal knowledge graphs.

Fully-inductive TKGR limitation: Existing RL-based multi-hop temporal knowledge graph reasoning methods perform poorly when training and test entities are entirely disjoint.

Sparse unseen-entity links: Sparse links for newly emerged unseen entities reduce available actions for multi-hop relational-path construction and impair reasoning accuracy.

TKG incompleteness: Temporal knowledge graph reasoning addresses the intrinsic incompleteness of temporal knowledge graphs.

Fully-inductive TKGR limitation: Existing RL-based multi-hop temporal knowledge graph reasoning methods perform poorly when training and test entities are entirely disjoint.

ARLIE: ARLIE is an adaptive reinforcement learning method with inductive embeddings for multi-hop reasoning in both fully-inductive and transductive temporal knowledge graph settings.

Context-based inductive embeddings: A context-based inductive representation method generates fine-grained embeddings for unseen entities using query-related contextual information.

Action-augmented adaptive RL: An action-augmented adaptive reinforcement learning framework leverages diverse actions to infer missing elements step by step over temporal knowledge graphs.

ARLIE: ARLIE is an adaptive reinforcement learning method with inductive embeddings for multi-hop reasoning in both fully-inductive and transductive temporal knowledge graph settings.

Context-based inductive embeddings: A context-based inductive representation method generates fine-grained embeddings for unseen entities using query-related contextual information.

SOTA TKGR improvement:

SOTA TKGR improvement: