Μια προσέγγιση μεταφοράς γνώσης σε συστήματα συστάσεων συνεργατικού φιλτραρίσματος με μη-επικαλυπτόμενες οντότητες
View/
Keywords
Transfer learning ; Recommender systems ; Collaborative filteringAbstract
Recommender systems are widely used in entertainment platforms and e-commerce businesses to personalize user experiences and enhance customer satisfaction. However, they face significant challenges, such as the sparsity problem, where interaction data is sparse, and the cold start problem, where new users or items lack sufficient interaction data for recommendation algorithms to train effectively.
Transfer learning is a machine learning approach that leverages knowledge gained from solving one problem to improve the solution of a different but related problem. Applying transfer learning techniques to recommender systems aims to address the challenges faced by traditional approaches, enabling the training of models with less data and improved performance.
This dissertation investigates the application of transfer learning in collaborative filtering recommender systems with non-overlapping entities. Specifically, the study focuses on scenarios where there are no common users or items between domains, using only user ratings to transfer knowledge without relying on metadata. The domains used are Movies and Books, and the experimental results are evaluated based on the accuracy of the proposed method compared to traditional recommendation techniques that do not utilize transfer learning.
Although the proposed algorithm demonstrated the potential to improve recommendations and positively transfer knowledge under certain conditions, it was not sufficient to outperform all traditional methods it was compared against, particularly the singular value decomposition (SVD) method.