A study of knowledge transfer techniques in recommender systems
View/
Keywords
Transfer learningAbstract
Recommender systems play a fundamental role in modern digital platforms by helping users discover relevant items among vast amounts of available content. Applications such as e-commerce, streaming services, and online media increasingly rely on recommendation algorithms to improve user experience and engagement. However, the effectiveness of traditional recommendation methods is often limited by data sparsity and cold-start problems, particularly when insufficient user–item interactions are available for learning reliable preference patterns.Transfer Learning has emerged as a promising approach for addressing these limitations by enabling knowledge acquired in one domain to be reused in another. In particular, Cross-Domain Recommender Systems aim to transfer useful information from a source domain to improve recommendation quality in a target domain. Most existing approaches, however, assume the existence of overlapping users or items between domains, an assumption that is often unrealistic in practical applications.
This thesis investigates a genre-based cross-domain recommendation framework designed for fully non-overlapping domains. The proposed approach uses the Book-Crossing dataset as the source domain and the MovieLens dataset as the target domain. Since the two domains share neither users nor items, knowledge transfer is achieved through semantic genre information. Book metadata are enriched using external sources, semantically aligned with movie genres using SentenceTransformer embeddings and cosine similarity, and subsequently integrated into a Genre-Aware Matrix Factorisation model.The framework introduces genre embeddings as shared semantic representations that capture relationships between user preferences and content categories. These embeddings are first learned in the source domain and then transferred to the target domain through pretrained initialization, enabling cross-domain knowledge transfer without requiring entity overlap. The transferred representations are subsequently fine-tuned on movie ratings and evaluated under progressively increasing sparsity conditions.Experimental evaluation is conducted using Singular Value Decomposition (SVD), Matrix Factorisation (MF), Genre-Aware Matrix Factorisation, and the proposed Cross-Domain Transfer Learning model. Performance is assessed using Mean Absolute Error (MAE) under 10-fold cross-validation and incremental training availability. The results demonstrate that incorporating semantic genre information significantly improves recommendation accuracy compared with traditional collaborative filtering methods. Furthermore, the findings indicate that semantic representations learned from an auxiliary domain can be successfully transferred to support recommendation tasks in a target domain.The proposed framework contributes to the development of more flexible recommendation systems capable of operating effectively under sparse-data conditions, while demonstrating the potential of semantic transfer learning in fully non-overlapping recommendation environments.
Postgraduate Studies Programme
Τεχνητή Νοημοσύνη - Artificial IntelligenceDepartment
Σχολή Τεχνολογιών Πληροφορικής και Επικοινωνιών. Τμήμα Ψηφιακών ΣυστημάτωνCorporate Department
National Center of Scientific Research "Demokritos"Number of pages
64Language
EnglishCollections
Except where otherwise noted, this item's license is described as
Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα
Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα