METHODS, TECHNOLOGIES, AND TOOLS OF INTELLIGENT INFORMATION PROCESSING PRODUCT RECOMMENDATION MODEL BASED ON DEEP NEURAL NETWORK ARCHITECTURE OF TRANSFORMER TYPE

This article considers the methodology for developing and the results of the experimental implementation of a product recommender system based on a deep neural transformer architecture with a mechanism of self-attention and the integration of additional user and item features. A detailed review of the evolution of recommender systems methods is provided: from classical approaches of collaborative and content-based filtering to modern hybrid and sequential models capable of capturing the dynamics of user interests over time. As the core, a modified transformer is used, which predicts the next user interaction taking into account the sequence of previous actions, the type of interaction, and various categorical and numerical characteristics of items and users. All features are encoded as trainable vectors, participating in the formation of the final sequence representation. Processing of long user sessions is implemented via a masked multi-head self-attention mechanism. The loss function is constructed as binary cross-entropy with negative sampling. A comparative analysis of the proposed model with a number of alternative methods (ALS, LightFM, ItemKNN, classical ranking models) was carried out using precision, recall, hitrate, and ranking quality metrics. It is shown that the proposed architecture provides a higher level of recommendation quality due to flexible integration of features and consideration of both short-term and long-term user behavioral dependencies. The experience of industrial deployment of the developed recommender system based on the Russian UX Rocket platform is described. Prospects for further development are discussed, including the integration of additional data sources and the optimization of computational.

Hali tarjima qilinmagan