Anchoring Path for Inductive Relation Prediction in Knowledge Graphs

Knowledge Graphs (KGs) are crucial for linking artificial intelligence (AI) with human understanding by modeling real-world and abstract concepts. However, due to their vast scope, KGs are often incomplete, necessitating relation prediction to infer missing triplets based on known KGs. Traditional relation prediction methods face two primary challenges: limited inductiveness and lack of explainability. Inductiveness is the ability to handle new entities as KGs grow, while explainability involves making predictions transparent and interpretable. Traditional methods, which assume static entities, struggle with new entities and typically provide predictions without insights. These challenges significantly hinder the effective application of KGs in real-world scenarios. To enhance the applicability of relation prediction in inductive and explainable contexts, this thesis proposes three novel approaches: Knowledge Reasoning Sentence Transformer (KRST), Anchoring Path Sentence Transformer (APST), and Context Pooling. KRST leverages pre-trained language models (PLMs) to encode reliable paths within KGs, introducing two innovative path extraction metrics (relation path coverage and relation path confidence) to filter out unreliable paths effectively. By modeling logical paths akin to Horn Rules, KRST provides accurate predictions and generates multi-perspective explanations derived from multiple relational paths. KRST represents the first Sentence Transformer model tailored for KG path encoding and demonstrates strong generalization across inductive and transductive settings. Experiments on benchmark datasets show that KRST outperforms SOTA models in 15 of 18 cases, establishing both its effectiveness and interpretability. Building on KRST, APST addresses the limitations of traditional path-based methods that rely solely on Closed Paths (CPs) in highly incomplete KGs. APST introduces the concept of Anchoring Paths (APs), enabling the model to utilize xiii xiv supporting evidence beyond CPs, thereby enhancing both prediction accuracy and explanation depth. APST unifies CPs, APs, and external textual knowledge extracted from public resources within a single Sentence Transformer framework. This integration yields substantial improvements in predictive performance and explanation quality. Evaluated on the same benchmarks, APST achieves SOTA results in 30 of 36 transductive, inductive, and few-shot scenarios, highlighting its robustness and adaptability. Furthermore, this thesis explores the enhancement of Graph Neural Network (GNN)based methods for explainable inductive relation prediction through Context Pooling. This innovative approach tackles neighbor selection challenges by identifying logically relevant neighbors for specific queries and applies graph pooling techniques within KGs to create query-specific graphs. Experimental results across multiple KG datasets show that Context Pooling achieves SOTA performance in 42 of 48 cases, while also enhancing model efficiency and scalability. In summary, this thesis advances relation prediction in KGs by enhancing inductiveness, explainability, and prediction performance. These methods strengthen trust and applicability in complex, high-stakes domains by integrating inductive reasoning and multi-aspect explainability.