Use of latent factors and consumption patterns for the construction of a recommender system

Abstract

In recent years, the recommender systems have become essential tools for companies to offer their products in a personalized way and to improve the user experience. The primary objective of these systems is to propose products or services to the user, according to specific criteria, such as their interests, their preferences, the place where they work or where they live. The problem arises when the system recommends products from an establishment to users who never visited that establishment. Besides, it is known that the order in which users purchase certain products or services can impact on the recommendation. To deal with these two problems, we propose a process that combines two widely used models: latent factors and matrix factorization. Also, to include temporality in our results, we use the \emph{Sequitur} algorithm. In order to test our proposal, we have used a database with approximately $65$ million banking transactions. The results obtained show the efficiency of our proposal in terms of percentage of success.