Abstract
Most of collaborative recommender systems (CRSs) rely on statistical and data analysis methods for comparing users. However, dealing with them using machine learning techniques seems to be more appropriate. This paper investigates the usage of feature selection and classification methods for CRSs. It suggests building a user model suitable for the classification purpose and proposes a density-based feature selection (DBFS) method based on the rating density for each class. The DBFS reduces the effect of sparsity problem and keeps only users having a dense-feature history. Additionally, a cascaded feature selection method is proposed to pick out a subset of features through a two-layer approach. The first layer applies a classical feature selection method while the second layer applied the DBFS on the output of the first layer. The results show that the performance is gradually improved. The cascaded feature selection yields the best results since it improves the system accuracy, reduces the space and processing complexities, and alleviates the sparsity in two cascaded layers. The achieved improvements by cascaded feature selection as compared to SVM are 6.55%, 10.14%, and 3.92% in terms of accuracy, F-measure and MAE, respectively.