Year: 2026
Type: article
Abstract: Educational recommender systems (ERS) contribute to the development of education and improve student engagement in the subject matter, as the system is tailored to personal preferences. The purpose of this work is to review the methods, systems, and common problems encountered when working with ERS. In particular, the Cold-Start problems in launching new items. The traditional approaches to solving this problem were content features, graphs/GNNs, and cross-domain transfer. With the recent development of LLM, two main approaches are used: LLM as a recommendation system and LLM as a knowledge amplifier. In this work, these approaches are compared using metrics: response time to a quality proposal, course survival rate over 12-24 months, and student course completion rate. Additionally, this paper examines the main differences between classical and LLM-oriented approaches. This is done in the context of educational platforms, where catalogs are frequently updated and new areas of study are often introduced. Particular attention is paid to three cold start scenarios: a new learner with no interaction history, a new course with a minimum number of clicks and reviews, and a new region. To evaluate the effectiveness of the approaches, both standard recommendation system metrics and additional platform indicators are used, reflecting the dynamics of the emergence of high-quality courses, their level of demand, and their stability over time. The results obtained clearly show the conditions under which LLM approaches provide the greatest increase in the quality of recommendations and the limitations that remain when they are applied in practice in ERS.
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
SDG: Quality education
Open Access status: bronze
OpenAlex ID: W7142930279