Regularizing Autoencoder-Based Matrix Completion Models via Manifold\n Learning

Autoencoders are popular among neural-network-based matrix completion models\ndue to their ability to retrieve potential latent factors from the partially\nobserved matrices. Nevertheless, when training data is scarce their performance\nis significantly degraded due to overfitting. In this paper, we mit- igate\noverfitting with a data-dependent regularization technique that relies on the\nprinciples of multi-task learning. Specifically, we propose an\nautoencoder-based matrix completion model that performs prediction of the\nunknown matrix values as a main task, and manifold learning as an auxiliary\ntask. The latter acts as an inductive bias, leading to solutions that\ngeneralize better. The proposed model outperforms the existing\nautoencoder-based models designed for matrix completion, achieving high\nreconstruction accuracy in well-known datasets.\n