Recommender Systems: New Algorithms and Current Practices

© iStock

The AI and Digital Science Institute at the HSE Faculty of Computer Science hosted a conference focused on cutting-edge recommender system technologies. In an atmosphere of active knowledge sharing among leading industry experts, participants were introduced to the latest advancements and practical solutions in recommender model development.

The conference brought together experts in the development of recommender systems—a promising technology with applications in both academia and industry. The conference was organised by the Laboratory for Matrix and Tensor Methods in Machine Learning headed by Maxim Rakhuba.

Evgeniy Frolov

According to Evgeniy Frolov, Senior Research Fellow at the laboratory and Head of the Personalisation Technologies Group at AIRI, 'The second iteration of the Recommender Systems Conference brought together a community of industry and academic experts, highlighting both a strong technological foundation and a growing interest in the field. The conference programme covered a wide range of topics, from recent research submitted to RecSys 2025—the leading conference on recommender systems—to in-depth reviews of production architectures used by major companies. A notable highlight was the roundtable discussion on how well-tuned single-stage solutions could serve as a stepping stone toward a unified, LLM-oriented recommender paradigm. From my perspective, the main outcome of the conference is the emergence of a community of industry and academic experts that enables honest hypothesis testing on real-world data and provides immediate insight into its value for both business and science.'

© HSE University

At a training seminar held during the conference, AI researchers Gleb Mezentsev and Danil Gusak provided a detailed overview of modern approaches to building scalable and consistent recommender systems. Participants explored the latest approaches to building efficient data pipelines for processing large amounts of data, as well as the complexities of integrating recommender solutions into real-world business processes.

Sergey Ermilov, Senior Developer at VK AI, presented research findings on the impact of advertising integrations on recommender service effectiveness and outlined successful strategies for content relevance and advertising returns.

Ruslan Israfilov, Sber RecSys Team Leader, delivered a presentation titled 'The Next Step in AI Evolution: LLM-based Multi-Agent Systems,' highlighting the benefits of integrating multiple intelligent agents to improve recommendation accuracy and better understand user behaviour.

Marina Ananyeva

Marina Ananyeva, Head of RecSys at the Laboratory for Matrix and Tensor Methods in Machine Learning, discussed the shift from traditional batch learning methods to online recommender models. She presented practical cases illustrating the transition to online learning, underscoring how models adapt more quickly to changes in audience preferences.

Alexey Vasilev, Executive Director of Data Science at the Sber AI Lab, emphasised the critical role of proper data preparation in developing high-quality recommender systems. His presentation covered topics such as model architecture selection, training process optimisation, and algorithm result interpretation. 'The conference was attended by experts from leading Russian companies. I know many of the speakers personally, so it was a pleasure to reconnect,' says Alexey Vasilev. 'The excellent variety of presentations—from both industry and academia—along with the poster session, made the event truly interesting. It was great to see the discussions continue during the breaks, once again confirming that recommender systems are a highly relevant and in-demand topic. I believe the conference was a success.'

In his presentation, Evgeniy Frolov proposed an innovative approach to enhancing recommender system performance by dynamically adjusting the structure of internal data representations—a method that can significantly improve recommendation quality and reduce the likelihood of errors. 'At the conference, I presented our new paper introducing a self-supervised approach to training recommender models. We adapted the Barlow Twins method, originally developed in the field of computer vision, for transformer-based recommender architectures. In particular, beyond improving prediction quality, we were the first to identify the effect of adaptive collapse in representations: depending on the structure of user preferences, the algorithm automatically adjusts the diversity of its outputs. In datasets without clear clusters of user tastes, it generates a broad range of recommendations, while in scenarios with strictly defined, specific preferences, it focuses on the most relevant products—delivering more accurate choices compared to existing methods,' explains Frolov.

© HSE University

The conference concluded with a poster session in the atrium of HSE University's building on Pokrovsky Bulvar, where participants discussed the presented research in an informal setting and exchanged ideas on emerging directions in recommender technology development.

This was the second Conference on Recommender Systems hosted by HSE University, and it is becoming a key platform for discussing scientific breakthroughs and technological innovations in AI and the digital economy. The event contributes to the advancement of the recommender systems industry and the emergence of a new generation of professionals in the field.