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.

48 training in distributed environments, and Non-negative Matrix Factorisation (NMF) is applied to interpretable recommendation models in cases involving non-negative data [4]. Despite its widespread adoption, collaborative filtering is associated with a number of engineering challenges. One of the key issues is data sparsity in interaction matrices, which arises from the large number of items and the limited number of user interactions, thereby negatively affecting prediction accuracy. In addition, the cold- start problem constrains the effectiveness of recommendations for new users or items due to the lack of historical data. The computational complexity of factorisation algorithms and the need to process large-scale datasets necessitate the use of distributed computing platforms, such as Apache Spark, Hadoop, and GPU-oriented environments [6, p. 15]. In e-commerce practice, collaborative filtering is used for the personalisation of product recommendations, the implementation of cross-selling mechanisms, and the creation of personalised product catalogues. In particular, recommendations such as ―Customers who bought this also bought‖ and ―Recommended for you‖ are based on item-based collaborative filtering algorithms, which enable an increase in average order value and user conversion rates [5]. In contrast to collaborative filtering, content-based recommendation algorithms are grounded in the analysis of item characteristics and generate recommendations based on the similarity between content features and the user preference profile. In e- commerce systems, such algorithms exploit a wide range of product features, including structured metadata, textual descriptions, categorical attributes, and visual features derived from product images [7]. A key stage in the development of content-based recommendations is the construction of vector representations of content. For textual data, term weighting methods such as TF–IDF are applied, along with distributed semantic representations of words and documents (word embeddings and sentence embeddings), which enable the modelling of latent semantic relationships between items. For product images, convolutional neural networks and pre-trained computer vision models are employed