A comparative study of regularised SVD and item-based kNN for movie recommender systems

Sammanfattning: This thesis compares the performance of two algorithms for rating predictions in movie recommender systems. The two algorithms examined, regularised singular value decomposition (RegSVD) and item-based k-Nearest Neighbour (item-based kNN), are compared on 9 different datasets. These datasets consists of constellations of 1,000-100,000 users and 100-10,000 movies. The problem statement is to find which algorithm performs the best on each dataset with respect to both accuracy and speed. These results are then  compared in order to identify general tendencies. The experiments are performed using the implementations of the algorithms in the LibRec library, where accuracy is measured using root-mean-square error (RMSE). Finally, the results show that item-based kNN outperforms RegSVD with respect to accuracy when evaluated on smaller datasets. However, RegSVD is a better alternative for larger datasets with respect to both accuracy and execution time.