Comparing Recommender Systems for a Large Assortment
In today’s online world where content is abundant and attention is scarce, showing the right products to a consumer is becoming increasingly important. Tools that address this problem have been researched intensively over the years. The release of a large consumer rating data set released by Netflix propelled research on this topic. In this thesis three different models for online recommendations are compared: Matrix Factorization, Non-Negative Matrix Factorization and the Cluster Affiliation Model for Big Networks. Using the BigCLAM algorithm as a recommender system is a novel approach since the algorithm is originally developed to be used as a community detection model. The models are applied to purchasing data from a large online retailer after which the results are compared in terms of recommendations, estimation speed and interpretability. The results show that the BigCLAM model is an interesting and effective alternative that may offer higher predictive performance that requires less estimation time.