DuET: Dual-View Tensor-to-Topology Spectral Adapter for Enhancing Sparse Tensor Factorization

Abstract

Many real-world datasets, ranging from web-interaction logs to biomedical networks, can be represented as sparse tensors where most entries are unobserved. Tensor factorization (TF) learns latent representations and a predictor to estimate unobserved entries and has been widely applied to higher-order recommendation, biomedical retrieval, and completion. However, under extreme sparsity, many entities participate in only a few interactions, yielding noisy and undertrained latent vectors with poor generalization. We propose DuET, a model-agnostic refinement framework that stabilizes TF embeddings by separating structural denoising from tensor factorization. DuET distills a dual-view structural prior from observed tuples and refines it via per-view low-rank spectral filtering and mode-wise gated residual refinement. We provide a theoretical guarantee that the resulting refinement is Frobenius-norm non-expansive, ensuring stability under repeated refinement. DuET seamlessly integrates with existing TF models without modifying their scoring architectures. Experiments on nine real-world tensors demonstrate consistent improvements for both retrieval and completion, with gains of up to 52.1% in NDCG@10 and reductions of up to 13.8% in RMSE.