Matrix factorization based heuristics for constraint-based recommenders

The main challenges for recommender systems are: producing high quality recommendations and performing many real-time recommendations per second for millions of customers and products. This paper addresses both challenges in the context of constraint-based recommenders where users specify their requirements and the system recommends a solution. We propose a novel approach to determine value ordering heuristics on the basis of matrix factorization. As far as we are aware, no researches exist in constraint-based recommendation domain which exploit matrix factorization techniques. The main idea of our approach consists in the prediction of value ordering heuristics based on historical transactions which can either represent past customer purchases or requirements. Thereby, value ordering heuristics are computed which are specific to each user's requirements. A series of experiments on real-world datasets for calculating constraint-based recommendations has shown that our approach outperforms compared methods in terms of runtime efficiency and prediction quality.