Regularizing autoencoder-based matrix completion models via manifold learning
Host Publication: European Signal Processing Conference
Authors: D. Nguyen, E. Tsiligianni, R. Calderbank and N. Deligiannis
Publication Year: 2018
Number of Pages: 5
Autoencoders are popular among neural-networkbasedmatrix completion models due to their ability to retrievepotential latent factors from the partially observed matrices.Nevertheless, when training data is scarce their performance issignificantly degraded due to overfitting. In this paper, we mitigateoverfitting with a data-dependent regularization techniquethat relies on the principles of multi-task learning. Specifically,we propose an autoencoder-based matrix completion model thatperforms prediction of the unknown matrix values as a maintask, and manifold learning as an auxiliary task. The latter actsas an inductive bias, leading to solutions that generalize better.The proposed model outperforms the existing autoencoder-basedmodels designed for matrix completion, achieving high reconstructionaccuracy in well-known datasets.