Architecture and optimization of the recommendation module of a higher education institution's electronic library under limited computational resources
DOI:
https://doi.org/10.34185/1562-9945-5-166-2026-16Keywords:
collaborative filtering algorithm, prediction, computational optimization, recommender system, temporal decay, digital library, MySQL, RMSEAbstract
The rapid digitalization of education inevitably causes information overload for students. While modern recommender systems rely on deep learning, these neural approaches are highly resource-intensive. They require GPU accelerators and extensive memory, making them unfeasible for local university servers typically operating on standard PHP/MySQL stacks. This paper presents the development and architectural optimization of a recommendation module tailored for higher education digital libraries functioning under strict computational constraints.
To address this, a user-based collaborative filtering method was applied, integrated with an exponential time decay function. This mathematical modification effectively mitigates "concept drift" by adapting the user profile to rapid shifts in academic interests across semesters, assigning significantly lower weights to outdated historical interactions.
To overcome the inherent quadratic computational complexity of classical algorithms, a highly optimized relational database architecture was engineered. The recommendation lifecycle was divided into two isolated phases. The first phase involves the asynchronous background recalculation of similarity matrices via queue managers, preventing real-time server overload. The second phase ensures the synchronous retrieval of pre-calculated suggestions from a denormalized data mart. By utilizing composite B-Tree indexing, the query response complexity for the end-user was reduced to a logarithmic level.
Empirical modeling on a real dataset (up to 1350 objects) confirmed the algorithm's high robustness. Experimental results proved that incorporating the time decay factor reduces the Root Mean Square Error (RMSE) from 2.10 to 1.21, an accuracy improvement of over 42%. The study substantiates that modified classical algorithms offer an optimal trade-off between prediction accuracy and computational cost for local educational platforms.
References
1. Tarasov, O., Vasylieva, L., Turlakova, S., Sahaida, P., & Karpov, A. (2023). Improving the quality of the relevance of the search for scientific publications based on a combination of ranking methods. In Proceedings of the 11-th International Conference "Information Control Systems & Technologies" (CEUR Workshop Proceedings, Vol. 3513). CEUR-WS.org. https://ceur-ws.org/Vol-3513/paper36.pdf
2. Mitreva, E., Paneva-Marinova, D., Georgiev, V., Nikolova, A., & Pavlov, R. (2026). A hy-brid approach for personalized and intelligent content recommendation in digital libraries. Applied Sciences, 16(6), Article 2756. https://doi.org/10.3390/app16062756
3. Tarasov, O., Sahaida, P., Podlesny, S., & Vasylieva, L. (2022). Categorical-ontological approach to information support of educational activities. SHS Web of Conferences, 142, Arti-cle 03005. https://doi.org/10.1051/shsconf/202214203005
4. Krishnamoorthi, S., & Shyam, G. K. (2023). Deep learning-based recommender systems—A systematic review and future perspective. In V. Bhateja, F. Carroll, J. M. R. S. Tavares, S. S. Sengar, & P. Peer (Eds.), Intelligent Data Engineering and Analytics. FICTA 2023 (Vol. 371). Springer. https://doi.org/10.1007/978-981-99-6706-3_33
5. Scrivano, A. (2025). A comparative study of recommender systems under Big Data con-straints [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2504.08457
6. Hartatik, Heryawan, L., & Pulungan, R. (2025). Trust decay-based temporal learning for dynamic recommender systems with concept drift adaptation. IEEE Access, 13, 110955-110971. https://doi.org/10.1109/ACCESS.2025.3583186
7. Ivokhin, Ye., & Sheliakin, H. (2024). Kolaboratyvna filtratsiia yak odyn iz osnovnykh me-todiv rekomendatsii kontentu: osnovni problemy ta sposoby yikh vyrishennia [Collaborative filtering as one of the main methods of content recommendation: main problems and ways to solve them]. Modern Engineering and Innovative Technologies, 1(34-01), 94–105. https://doi.org/10.30890/2567-5273.2024-34-00-013 [in Ukrainian]
8. Wang, Y. (2025). Time-weighted collaborative filtering recommendation algorithm based on large language models. In P. T. Mathiopoulos & Y. Feng (Eds.), Sixth International Con-ference on Signal Processing and Computer Science (SPCS 2025) (p. 59). https://doi.org/10.1117/12.3089553
9. Khabarlak, K., & Shyriaiev, A. (2025). Porivnialnyi analiz modelei mashynnoho navchan-nia dlia pobudovy systemy rekomendatsii z dynamichnym perenavchanniam [Comparative analysis of machine learning models for building a recommendation system with dynamic re-training]. Information Technology: Computer Science, Software Engineering and Cyber Secu-rity, (4), 267–273. https://doi.org/10.32782/it/2025-4-31 [in Ukrainian]
10. Singhal, A. (2001). Modern information retrieval: A brief overview. Bulletin of the IEEE Computer Society Technical Committee on Data Engineering, 24(4), 35–43. https://www.cs.columbia.edu/~gravano/Qual/Papers/singhal.pdf
11. Alpert, S. I. (2019). Osnovni miry podibnosti ta novi pidkhody do yikh zastosuvannia pry klasyfikuvanni hiperspektralnykh kosmichnykh zobrazhen [Basic similarity measures and new approaches to their application in classification of hyperspectral space images]. Mate-matychni Mashyny i Systemy, (1), 143–151. https://www.immsp.kiev.ua/publications/articles/2019/2019_1/01_2019_Alpert.pdf [in Ukrainian]
12. Novotný, V. (2018). Implementation notes for the soft cosine measure. In The 27th ACM International Conference on Information and Knowledge Management (pp. 1639–1642). As-sociation for Computing Machinery. https://doi.org/10.1145/3269206.3269317
Downloads
Published
Issue
Section
License
Copyright (c) 2026 System technologies
This work is licensed under a Creative Commons Attribution 4.0 International License.
