Multicriteria evaluation of recommender systems using fuzzy analytic hierarchy process

Abstract

Offering relevant items and goods has always been a core issue for consumer-centric business. In recent years the digitalization process gave such entities a significant push towards implementing computer-aided systems for solving recommendations generation problems via applying statistical learning methods on large scale datasets of user-item interactions. Since then, there were developed a variety of algorithms for a recommender system design, based on neighborhood metrics, matrix factorization machines and deep neural networks, each of them having their application criteria and requirements on input data. The purpose of the following research is to investigate performance of graph neural networks recommender systems on two tasks – recommendation retrieval and next best offer prediction. As a result, the graph-based recommender system for solving both tasks, described above, was synthesized and benchmarked on open-source e-commerce dataset. From the conducted experiment it was gained evidence that proposed approach is feasible to solve recommendation retrieval tasks and outperform models, that are considered de-facto standard, i.e., factorization machines and recurrent neural netwoks.