Imputation strategies for cold-start users in NMF-based recommendation systems

We propose a recommendation system method which is based on NMF (Nonnegative Matrix Factorization) in collaborative filtering to enhance the rating predictions. The proposed method conduct selective imputations that fuses the factored original rating matrix and the factored imputed rating matrix into one system. The outputs of the factorized matrices provide four different ways to calculate the predicted ratings which are called sub-predicted ratings. Our proposed method is capable of predicting the rating by utilizing either the imputed users, or imputed items, or both in order to limit the errors that may be introduced from the imputed ratings. We proposed five strategies to calculate the final predicted rating from the sub-predicted ratings. The prediction results of rating values that are not close to the average of the rating values could be enhanced by utilizing the proposed method. Experiments on four different datasets are conducted to examine the proposed approach. The results show that our approach improves the predicted rating especially with Max of value category strategy.