FT-MixE: Fourier transform-based efficient mixing of knowledge graph embeddings for improved link prediction

Knowledge graphs (KGs) are inherently incomplete due to the vast and continuously evolving nature of real-world knowledge. KG completion (KGC) addresses this limitation by predicting missing links, which requires learning expressive representations of semantic relationships. Concurrently, computational efficiency is critical for scal ing KGC models to large real-world graphs. However, existing approaches struggle to balance expressiveness and efficiency, as richer semantic embeddings often demand higher-dimensional representations. We introduce a novel KGC model, the Fourier Transform-based Efficient Mixing of KG Embeddings for Link Prediction (FT-MixE). Our method leverages the two-dimensional discrete Fourier transform (DFT), a powerful tool for mixing token embeddings in the realm of natural language processing. FT-MixE applies the non-parametric DFT to de sign an encoder that effectively mixes entity and relation representations, achieving both accuracy and efficiency in the KGC task. Our model demonstrates competitive link prediction performance with superior parameter ef ficiency. Specifically, FT-MixE achieves higher MRR and Hit@k (k = 1, 3, 10) on FB15k-237 and YAGO3-10, and comparable or better performance on WN18RR, relative to strong baselines, while using substantially fewer parameters. ; Y ; 1