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.

49 to generate image embeddings that capture the visual properties of products in a low- dimensional vector space [8]. The engineering implementation of content-based recommendations involves constructing a user profile by aggregating the vector representations of items with which the user has interacted. Such a profile may be formed as a weighted average vector of the features of viewed or purchased items, or as a more complex model employing neural networks and attention mechanisms to capture the dynamics of user preferences. In e-commerce practice, content-based models are integrated into catalogue personalisation systems, similar-item recommendation modules, and search mechanisms with semantic relevance [6]. At the same time, content-based recommendation algorithms exhibit several limitations. One of the key challenges is the informational sparsity of the user profile, which arises from a limited number of interactions or insufficient quality of product descriptions. In addition, such models are prone to the filter bubble effect, whereby recommendations are confined to a narrow set of similar items and fail to provide sufficient diversity, which may negatively affect user experience and the commercial performance of the platform [3, p. 50]. To overcome the limitations of collaborative and content-based algorithms, hybrid recommender models are widely employed in modern e-commerce systems, as they combine multiple data sources and algorithmic approaches within a unified architecture [1, p. 10882]. Hybrid systems enable the integration of user behavioural patterns with content characteristics of items, thereby improving recommendation accuracy and model stability in dynamic e-commerce environments. The academic literature identifies several architectural approaches to combining recommendation methods. The feature-level fusion approach involves integrating features from collaborative and content-based models at the input data level, followed by training a unified model [3, p. 51]. Model-level fusion is based on combining the outputs of multiple independent models, for example through linear combination or neural aggregation of their predictions. The ensemble approach employs ensemble