JOINT FEATURES-BASED KNOWLEDGE GRAPH COMPLETION | ||
| International Journal of Intelligent Computing and Information Sciences | ||
| Volume 25, Issue 2, June 2025, Pages 88-104 PDF (728.82 K) | ||
| Document Type: Original Article | ||
| DOI: 10.21608/ijicis.2025.398811.1407 | ||
| Authors | ||
| Maha Farghaly* 1; Mostafa Aref2; Mahmoud Mounir3 | ||
| 1Information Systems Department, Faculty of Computers and Information Technology, Future University in Egypt, Cairo, Egypt. | ||
| 2Department Computer Science, Faculty of Computer and Information Sciences,Ain Shams University, Cairo, Egypt. | ||
| 3Information Systems Department, Faculty of Computer and Information Sciences, Ain Shams University, Cairo, Egypt | ||
| Abstract | ||
| Knowledge graphs (KGs) help in resolving data inconsistencies and redundancies by organizing information in a unified structure, paving the way for building scalable, interpretable AI systems, as they provide a transparent way to trace reasoning paths and conclusions. The continuous evolving nature of information imposes for continuous knowledge graph completion (KGC). Thus, many research efforts are directed into building efficient KGC models. KGC models varied in the utilized knowledge features into the knowledge textual semantics, the graph relational structure, and temporal features, proving the crucial role of each. However, prior models neglect utilizing one or more types of features in the embedding and completion learning, missing the potential advantage of the neglected features for efficient completion learning. Thus, we propose an innovative sequential hybridization-based completion model, named Joint Features-based Knowledge Graph Completion (JF-KGC). The model encodes the latent semantics of knowledge, the structural topology, and the evolution of knowledge over time. This is a novel completion model that utilizes hybrid embedding of deep textual-based semantics, graph structure and temporal features for completion learning. The experiment conducted over YAGO11k and Wikidata12k benchmark datasets proved significant improvement over baselines in predicting missing knowledge with Mean Reciprocal Rank (MRR) equals 39.2% and 51.3% on YAGO11k and Wikidata12k, respectively. Besides, we provide an ablation study to prove and quantify the contribution of each feature type to the performance. | ||
| Keywords | ||
| Knowledge Graph Completion (KGC); Temporal Knowledge Graph (TKG); Knowledge Semantics; Graph Topology; Language Model (LM) | ||
|
Statistics Article View: 81 PDF Download: 71 |
||
