SG-NTF: New tensor method improves completion with spectral gating

Researchers Fusheng Wang and Yikai Hou have introduced Spectra-Guided Neural Tucker Factorization (SG-NTF), a novel approach for completing high-dimensional and incomplete (HDI) tensors. Traditional tensor completion methods struggle with discrete representations and fail to capture continuous temporal patterns. SG-NTF overcomes this by embedding scalar timestamps into a continuous spectral space, allowing the model to abstract temporal periodicities without predefined bins. Additionally, the Spatio-Temporal Co-Gating (STCG) mechanism explicitly filters latent interactions using multiplicative modulation on spatiotemporal contexts, enabling the model to focus on the most relevant dependencies.

Experiments on real-world HDI tensors demonstrate that SG-NTF maintains competitive completion accuracy while being parameter-efficient. This is especially valuable for applications like recommender systems, traffic forecasting, and climate modeling where data is often sparse and high-dimensional. By combining neural Tucker factorization with spectral guidance and co-gating, SG-NTF sets a new baseline for tensor completion tasks. The paper is available on arXiv (2606.00584) and represents a significant step in bridging continuous representation learning with tensor algebra.