Comparing neighborhood and matrix factorization models in recommendation systems : Saving the user some clicks

Abstract: This thesis explores how different recommendation models based on machine learning can be implemented using customer activity data from an internet portal. The recommendation models have been evaluated on both error metrics and accuracy metrics. In addition, two data-sets are used, one open and widely available called MovieLens, and one proprietary from the principal of this thesis. Furthermore, the models evaluated are based on popular neighborhood methods and more advanced matrix factorization techniques. Based on both error and accuracy metrics, the matrix factorization models SVD and SVD++ outperform the neighborhood model in both data-sets. Moreover, the SVD++ algorithm performed the best of all the models but in the end, SVD was chosen for the principal. This was done because it proved to be a better overall choice when looking at all the parameters such as performance, speed and ease of use.