Sber Researchers Shine as Top Russian Team in RecSys Challenge 2025

A team of researchers from Sberbank’s Practical Artificial Intelligence Center, known as ai_lab_recsys, achieved remarkable success at the international competition, the [RecSys Challenge 2025](https://recsys.synerise.com/). They secured the top position among academic teams and finished second overall, marking them as the best Russian team in this annual contest focused on recommendation systems. The competition featured over 100 teams and 400 experts from various countries.

Participants were tasked with creating universal behavioral profiles based on user actions—ranging from browsing and search queries to purchases and cart operations. The resulting vector representations (embeddings) needed to demonstrate strong generalization capabilities, with their quality assessed across six different tasks, including user churn prediction and forecasting interests in popular categories and products. The challenge was further complicated by the presence of hidden tasks with undisclosed criteria.

The ai_lab_recsys team employed a range of approaches to build the embeddings. In their final solution, they utilized a combination (concatenation) of different types of representations, from straightforward yet effective models such as ALS (Alternating Least Squares) and sets of user features to more complex architectures like autoencoders and Gated Recurrent Units (GRU). This strategy enabled them to achieve high levels of generalization and consistent results in both the open and hidden tasks of the competition.

Despite multiple changes to deadlines and dataset replacements, the team delivered a powerful and resilient solution that outperformed leading academic and industrial teams. The researchers at Sber are now preparing a scientific paper detailing their methods and findings to present at the official workshop of the RecSys’25 conference in Prague in the fall of 2025. The insights gained from this competition will also be applied to solve challenges faced by Sberbank.

*RecSys Challenge is an annual machine learning competition held in conjunction with the major international conference ACM RecSys. It is regarded as one of the most significant contests in the field of recommendation technologies due to its scientific depth and high level of competition.*