Neural Tucker Factorization

Abstract

This letter presents a novel latent factorization model for high dimensional and incomplete (HDI) tensor, namely the neural Tucker factorization (NeuTucF), which is a generic neural network-based latent-factorization-of-tensors model under the Tucker decomposition framework. It first interprets the traditional Tucker framework into a neural network with embeddings for different tensor modes. Afterwards, a Tucker interaction layer is innovatively built to accurately represent the complex spatiotemporal feature interactions among different tensor modes. Experiments on real-world datasets demonstrate that the proposed NeuTucF model significantly outperforms several state-of-the-art models in terms of estimation accuracy to missing data in an HDI tensor, owing to its ability of accurately representing an HDI tensor via modeling the complex interaction among different input modes. Interestingly, the results also indicate that our model has a certain level of implicit regularization.