Theoretical and Practical Aspects Building Recommendation Models: Typology, Architecture and Directions Design

Abstract

The article is devoted to the study of promising areas of designing a recommender system for the pre-order and delivery service RapiDo with an emphasis on deep learning methods and problems of models at a cold start. The authors analyze existing types of recommender systems, their features in pre-order and delivery services and aspects of insufficient efficiency of modern models associated with the lack of consideration of the order context, individual user preferences, irrelevance of the data used and the lack of feedback. The article considers the main types of recommender systems used by the largest Russian delivery platforms and user information that services use when building their recommender models, and also highlights the key areas of designing the RapiDo recommender system, taking into account the need to work with limited data at early stages. The authors pay special attention to the architectures of recommender models based on deep learning methods, considering more than a dozen of the most popular options. Promising approaches are analyzed, including adaptive learning based on user feedback, collaborative filtering using demographic data, and hybrid mechanisms that combine different methods to improve the accuracy and stability of predictions. The paper presents the first results of the study and highlights the importance of integrating deep learning into the RapiDo recommendation system to achieve high model accuracy and address the problem of insufficient data at early stages.