A recommender system suggests items that the user of the system is likely to find valuable. Together with the explosion of e-commerce, recommender systems have become a focus of academic research. Within this field prediction of film ratings is a popular research area and the topic of this thesis. Many websites that sell, rent or stream films allow users to give ratings to films they have seen. The goal is to accurately predict ratings the user has not given yet. It would then be possible to recommend films the user may want to see.
Different ways of predicting film ratings in recommender systems were compared at this thesis using MovieLens 100K as the dataset. The algorithms were implemented in MATLAB, tested using 5-fold cross-validation, and ranked using mean absolute error as the accuracy metric. In total nine different recommender system designs were tested, including four hybrid systems designed and created for this thesis. Techniques used include user and item-based collaborative filtering, singular value decomposition, content-based recommendation and demographic method. Separate tuning data was used to optimise parameters including similarity measure used and the best nearest neighbourhood size.
Of the basic methods item-based collaborative filtering gave the best results, followed by singular value decomposition. User-based collaborative filtering, content-based recommendation and demographic method performed slightly worse. The overall best results were achieved with a hybrid design that combines baseline predictors with user-based and item-based collaborative filtering. Choosing the best similarity measure and finding ideal values for parameters like nearest neighbourhood size had a significant impact on the results.
