Onto Proximality in Non Negative Matrix Factorization for Recommender Systems

Abstract

Recommender Systems have become integral to most e-commerce applications and online platforms. The recommended suggestions heavily impact customer retention and business performance. One of the essential parameters in large-scale recommender systems is the time required to present a recommendation. The more time it takes, the more it loses the customer’s attention and interest. It is currently necessary for recommenders to be time-efficient and optimal. Collaborative filtering-based Matrix Factorization approaches have proven to be powerful for recommender systems. The standard approach uses the Singular Value Decomposition-based recommender systems with Gradient Descent optimizer and its advanced variants. These models provide good accuracy for recommenders. However, they are time-intensive. To alleviate these issues, the proximal gradient descent optimizer-based Nonnegative Matrix Factorization model is adapted for recommender systems to improve their performance in terms of time and accuracy. There has been no research on integrating proximal descent models in Nonnegative matrix factorization for recommender systems. These novel adaptations are analyzed with six other baseline recommender models on two datasets. The experimental analysis proves that these novel recommender models are the preferable choice for online recommenders and work well when the data is not smooth.

Keywords

• Lipschitz continuity
• nonnegative matrix factorization
• recommender systems
• singular value decomposition
• stochastic proximal gradient descent