Movie recommendations using machine learning algorithms

Abstract

This research attempts to evaluate machine learning technology in a movie recommendation scenario. Previous research in the area has mostly used the MovieLens100K dataset and Mean Absolute Error (MAE) accuracy calculation mechanism; hence for comparison purposes, this research will apply these in its case studies. A review of previous literature shows that good accuracies could be obtained using various methods, however details about internal workings or execution speed are not always given. For this reason, standard machine learning technologies identified via a literature review have been individually examined in a consistent way that would then allow a fair comparison of their accuracy and performance. Furthermore, this research also proposes two novel machine learning technologies, specifically designed for Movie Recommendation. The Matrix technology can rate any kind of movie, even those that it has never encountered in its training while achieving decent accuracy and execution speed. The Movie Centric technology, on the other hand, concentrates on movies it has been trained on and generalises its viewers, achieving a better performance than the Matrix one both in terms of speed as well as MAE. Both technologies can work with just the viewer’s gender as opposed to the group of viewer parameters utilised by other standard technologies. This research has concluded that the Random Forest Classifier provides the best MAE/speed compromise between the standard technologies and the Gradient Boosting Classifier provides the best MAE at the expense of speed. The novel Movie Centric algorithm proposed outperforms the Random Forest Classifier both in speed as well as MAE using only viewer gender. However, it does not reach the accuracy levels of the Gradient Boosting Classifier although it executes much faster.