Invertible Neural Networks for Matrix Factorization in Recommendation Systems

Resumen

Recommendation systems play a central role in personalizing digital content by leveraging user–item interactions. Matrix factorization is widely used but has limitations, particularly when integrating both explicit feedback (ratings) and implicit feedback (clicks or lack of interactions).
In this work, we propose a framework based on invertible neural networks for matrix factorization. The interaction matrix A (M x N) contains explicit ratings and implicit feedback for M users and N items. The goal is to estimate missing entries to predict preferences and generate personalized recommendations.
Unlike classical methods based on dot products of latent vectors, our approach projects users and items into a low-dimensional latent space via a nonlinear and invertible transformation. This design preserves information, reconstructs the original interactions, and simultaneously exploits explicit and implicit feedback.
Experiments on benchmark datasets demonstrate the potential of the proposed model. The framework also applies to other loss functions and can integrate heterogeneous data sources (text, social networks, browsing history) in order to improve robustness and personalization.