Extending hypernetwork-based recommender systems for the cold-start problem using side information

Panis, Konstantinos; Πάνης, Κωνσταντίνος

Master Thesis

Author

Panis, Konstantinos

Πάνης, Κωνσταντίνος

Date

2026-03

Abstract

The cold-start problem constitutes a fundamental challenge in recommender systems, as it requires generating accurate personalized recommendations for users with limited or even no prior interaction history. Although the integration of side information has been widely adopted in traditional recommender systems, its effectiveness within meta-learning hypernetwork based framework remains insufficiently explored. This thesis investigates the capabilities of the HyperRS meta-learning model when heterogeneous side information is incorporated under different user cold start settings. Specifically, the integration of two types of side information is examined: NLP-based item representations derived from SBERT-encoded movie plots and user demographic data. Three extensions of the original HyperRS model are proposed, integrating SBERT-derived movie-plot embeddings, user demographic information and their combination, respectively. The experimental results indicate that semantically enriched item side information provides marginal performance improvements under moderate cold-start conditions but becomes less effective as cold-start severity increases. In contrast, user demographic information appears to contribute weak personalization signal and does not consistently improve performance. Overall, the results demonstrate that the effectiveness of additional side information in hypernetwork based recommender systems is highly dependent on interaction data sufficiency and that interaction-based signals remain fundamental even within meta-learning frame works. The main contribution of this study lies in the systematic extension and evaluation of hypernetwork-based reccomender systems through the integration of different types of side information under varying cold-start conditions, highlighting important considerations for the integration of heterogeneous side information in practical recommendation scenarios.

Postgraduate Studies Programme

Πληροφοριακά Συστήματα και Υπηρεσίες

Department

Σχολή Τεχνολογιών Πληροφορικής και Επικοινωνιών. Τμήμα Ψηφιακών Συστημάτων

Number of pages

Language

English

URI

https://dione.lib.unipi.gr/xmlui/handle/unipi/19073

Collections

Τμήμα Ψηφιακών Συστημάτων

Show full item record

Except where otherwise noted, this item's license is described as
Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα