Multi-Feature-based Knowledge Graph Completion: Features Attention vs. Score Ensemble for Feature Aggregation

Abstract

The incompleteness of knowledge graphs obstructs the efficiency of relying-applications. Knowledge Graph Completion (KGC) leverages knowledge features to learn completion, which can be categorized into semantics, topologies, and temporals. Despite the proven advantage of each type over completion learning, KGC models utilized single or dual feature types, missing the potential advantage of the neglected type. We aim to maximize features utility by incorporating the three types to advance KGC and improve the generalization and reliability in case of inadequacy or unreliability of specific features in the graph. We propose Multi-Feature-based Knowledge Graph Completion (MF-KGC) which learns completion with the incorporation of semantics guided by large language model, topologies, and temporals. Two versions of MF-KGC are proposed based on the features aggregation approach: Attention-based MF-KGC (AMF-KGC) and Ensemble-based MF-KGC(EMF-KGC). Evaluation demonstrates that AMF-KGC achieved 50.6% MRR, 45.0% Hits@1, 57.4% Hits@3, and 69.0% Hits@10 on Wikidata12k and 38.1% MRR, 33.3% Hits@1, 41.5% Hits@3, and 59.7% Hits@10 on YAGO11k, while EMF-KGC achieved 49.8% MRR, 41.7% Hits@1, 55.7% Hits@3, and 68.2% Hits@10 on Wikidata12k and 37.6% MRR, 28.7% Hits@1, 40.0% Hits@3, and 55.2% Hits@10 on YAGO11k, proving the superiority of both versions over competitors. Investigative feature-elimination study was performed to quantify the impact of each.