Few-Shot Fuzzy Temporal Knowledge Graph Completion via Fuzzy Semantics and Dynamic Attention Network

Few-shot temporal knowledge graph completion (TKGC) aims to predict missing facts in temporal knowledge graphs (TKGs) when each relation has a limited number of available samples. Existing few-shot TKGC models primarily face two challenges. First, entities and relations in TKGs typically exhibit dynamic characteristics, and those models often adopt static neighbor information to improve their semantic representations, so the noise in neighbor information may be amplified. Second, those models mainly focus on crisp entities and relations, yet real-world facts are often fuzzy, rendering them ineffective at capturing fuzzy semantic information. To solve these challenges, this article investigates a novel method for modeling fuzzy TKGs information in a vector space and proposes a framework for few-shot fuzzy temporal knowledge graph completion. Specifically, we devise a dynamic attention neighbor aggregator and time-aware fuzzy mapping framework. These components facilitate the integration of semantic information from neighborhood relations and generate fuzzy membership degrees to describe the fuzzy information. Additionally, we integrate an error-aware feature extractor within the meta-learning framework, effectively addressing errors during the learning process. Extensive experiments show that our model achieves outstanding results across different few-shot settings, obtaining performance gains of 4.2%, 13.6% in terms of MRR and Hits@10, respectively, on the ICEWS0515-ff dataset, and 2.9%, 10.2% on the ICEWS18-ff dataset.