Recommender System Optimization through Collaborative Filtering

The amount of available data is growing on a rapid pace. Due to this, data handling is becoming increasingly important. Netflix organised an online competition in 2006 to improve their recommender system. Recommender systems based on collaborative filtering optimize product suggestions based on user preferences by analyzing past user and item rating behaviour. The aim of this research is to optimize movie suggestions by predicting movie ratings based on past ratings. This is done by implementing the neighbourhood based collaborative filtering method described in Bell and Koren (2007), which uses the training data of the Netflix prize competition to train their model. This method differentiates itself by normalizing the data by correcting the ratings for commonly known item and user characteristics. Furthermore, it comes up with an approach which simultaneously derives the interpolation weights for all nearest neighbours. This improves the accuracy of the model in comparison with separately derived interpolation weights. This study adds an extension to this work by correcting for possible noise in the data. This is done in the weight selection, where high interpolation weights are becoming relatively more important than small interpolation weights. Furthermore, the hypothesis that frequently seen movies contain less valuable prediction information is investigated as a second extension. The results show that the addition of both extensions do not improve the prediction performance of the model.