Proceedings of the International scientific and practical conference ―Modern Science: Challenges and Perspectives‖ (February 9-11, 2026) / Publisher website: www.naukainfo.com. - London, United Kingdom, 2026. - 121 p.

47 neural network models are employed to represent complex multidimensional data structures [4]. From an architectural perspective, recommender systems are classified into offline batch learning models, online streaming recommendation systems, and hybrid architectures that combine batch and stream processing. In contemporary e- commerce platforms, recommendation algorithms are integrated into distributed microservices architectures using cloud computing and MLOps infrastructure, which ensures scalability, low-latency inference, and adaptive model updates [5]. Thus, the classification of recommendation algorithms by data sources, mathematical models, and architectural characteristics enables the development of a systematic engineering approach to their design and deployment in e-commerce. Collaborative filtering is one of the fundamental and most widely used approaches to building recommender systems in e-commerce, as it is based on the analysis of user behavioural patterns and the assumption of similarity in user preferences or item characteristics [6]. The core idea of this method lies in exploiting the history of user–item interactions to predict future preferences without the need to analyse the content characteristics of the recommended items. Within collaborative filtering, user-based and item-based approaches are distinguished. User-based collaborative filtering generates recommendations based on similarities between users, employing similarity metrics such as cosine similarity and Pearson correlation to identify groups of users with similar behavioural profiles. In contrast, the item-based approach relies on analysing similarities between items and generates recommendations by identifying objects similar to those with which a user has already interacted, thereby providing greater stability of results in large-scale e-commerce catalogues [7]. For scalable recommender systems, matrix factorisation methods have been widely adopted, as they enable dimensionality reduction of the user–item interaction matrix and the identification of latent preference factors. Singular Value Decomposition (SVD) is used to decompose the rating matrix into latent user and item matrices, Alternating Least Squares (ALS) enables efficient parallel model