![]() | Only 14 pages are availabe for public view |
Abstract Automated recommender/recommendation systems make items suggestions that are tailored to the users’ needs and represent powerful against big data information systems. However, their practical applicability still faces many problems compared with classification models. In this thesis, we started by defining the framework of the recommendation systems and how to build it using multi-label classification systems. Due to the low accuracy of the recommender system, information fusion method in different levels is proposed. These levels including feature level, decision level, rank level and score level. More details of each level are presented with examples. Moreover in this thesis, a novel ranking aggregation method was proposed to address the drawbacks of the Condorcet and non-Condorcet methods. Different experiments were conducted to compare the accuracy of recommender systems in different levels. These results proved that score, rank, and feature levels achieved high performance because both levels preserve useful information, while the decision level neglects more data, which may degrade the performance of recommender systems. The imbalanced dataset is one of the main problems in classification and recommendation systems. Many sampling methods were used to obtain a balanced dataset. Different assessment methods that are suitable for imbalanced datasets are used. Different experiments were conducted using imbalanced datasets and the results proved that the sampling methods that used improved the performance of recommendation systems. |