Tensor Methods Outperform Transformers: A Comprehensive Comparative Analysis of Knowledge Graph Link Prediction Methods

Knowledge graph link prediction has witnessed significant methodological evolution, with recent research heavily favoring transformer-based approaches. However, our comprehensive comparative analysis reveals surprising results: classical tensor factorization methods significantly outperform modern transformer architectures across multiple domains. We systematically evaluate four major categories of link prediction methods—classical embeddings (TransE, DistMult, ComplEx), tensor factorisation (CP, Tucker Decomposition), graph neural networks, and transformer approaches across three diverse knowledge graph datasets (UMLS, FB15K, DB15K). Contrary to prevailing trends, tensor methods achieved superior accuracy (Tensor-CP MRR: 0.2273) while transformer approaches demonstrated unexpectedly poor performance (LP-BERT MRR: 0.0236). Our multi-dimensional evaluation encompassing accuracy, computational efficiency, and scalability identifies three
distinct operational tiers and provides evidence-based guidelines for method selection. These findings challenge the current research focus on transformer architectures and suggest a renewed importance for tensor-based approaches in knowledge graph completion tasks