Temporal Matrix Factorization: A polynomial approach to latent factor estimation

Abstract

Matrix Factorization (MF) is a simple and efficient Machine Learning (ML) technique to discover latent factors that help in explaining the underlying behaviour of actors (for instance, in the domain of recommender systems the actors could be users/items). The technique uses observed data, generated by the actors, to derive the latent factors. With time, more and more observed data that spans across a wide range of time period gets accumulated. The abundance of this data, allows us to discover new behavioural patterns. One such pattern is that, the latent factor values change with time. Most of the existing matrix factorization techniques are agnostic to time of occurrence of the event and fail to foresee the temporal changes in the actor behaviour. Deep neural networks are effective in predicting the recommendations with time, however, the ability to explain the underlying pattern and the associated rationale is poor. In our work, we propose a novel matrix factorization technique to estimate the temporal changes in the user behaviour and demonstrate its effectiveness in achieving time dependent recommendations.