Multi-Level Ensemble Learning based Recommender System

In this paper, we have tried to go beyond conventional ensemble learning & explore multi-level ensemble learning with reference to recommender systems. In particular, we have focused on stacked generalization for building Movie Recommender System. We have tried to analyze the transition from single level to multi-level ensemble learning and its effects on the overall accuracy. We have used movielens dataset from Grouplens Project and used a host of techniques like Collaborative Filtering, PAM, Content based recommender, Random Forest, SVM, ANN, etc. to optimize accuracy. We have experimented with various combinations of base learners based on their accuracy & diversity to finally arrive at the most accurate ensemble of ensembles. Results show that 2-level stacking gives more accurate results than single level stacking or any individual recommender system.