Goran Gligorin (2011) Overview and comparison of recommender systems. EngD thesis.
The field of recommender systems is most commonly classified into three main categories: content--based, collaborative filtering and hybrid recommendation algorithms which combine the features from the first two categories. The goal of the thesis was the implementation of two algorithms from the collaborative category, today the most commonly used basis for recommender systems, and their evaluation. Collaborative filtering is further divided into two groups: memory--based and model--based methods. From implementation we chose one algorithm form each group. We chose a neighborhood--based method and a method based on matrix factorization to represent each of the groups respectively. We implemented an extra method that combines the properties of the first two. The results of testing showed that building a recommender system that performs better then naive methods. The analysis showed that the main reasons lie in data sparsity problem, which is one of the main problems collaborative filtering methods face. As expected matrix factorization, which is designed to handle this problem, produced better results than other methods. In the conclusion we present some ideas for further work, which include estimate calibration and excluding unrepresentative artists.
Actions (login required)