Improving the scalability of recommender systems by clustering using genetic algorithms

Abstract

It is on human nature to seek for recommendation before any purchase or service request. This trend increases along with the enormous information, products and services evolution, and becomes more and more challenging to create robust, and scalable recommender systems that are able to perform in real time. A popular approach for increasing the scalability and decreasing the time complexity of recommender systems, involves user clustering, based on their profiles and similarities. Cluster representatives make successful recommendations for the other cluster members; this way the complexity of recommendation depends only on cluster size. Although classic clustering methods have been often used, the requirements of user clustering in recommender systems, are quite different from the typical ones. In particular, there is no reason to create disjoint clusters or to enforce the partitioning of all the data. In order to eliminate these issues we propose a data clustering method that is based on genetic algorithms. We show, based on findings, that this method is faster and more accurate than classic clustering schemes. The use of clusters created, based on the proposed method, leads to significantly better recommendation quality