Researchers at HSE University and Sber Train Neural Networks to Better Predict User Preferences

© iStock

The HSE FCS AI and Digital Science Institute and Sber have introduced a new architecture for recommendation systems that combines two classes of models, enabling algorithms to better predict users’ interests and needs. A preprint of the paper has been published on arxiv.org and presented at Urban ML.

Recommendation systems are programs and algorithms that analyse users’ behaviour, preferences, and interaction history to predict which products or services they are most likely to be interested in. Modern recommendation systems are based on two main classes of models: transformers and graphs. Transformers are effective at predicting the sequence of products in a user’s history but do not capture global connections between users and products. Graph-based models, by contrast, are well suited for modelling such connections but do not account for temporal dynamics as effectively. As a result, they are less accurate than transformers at predicting a user’s next choice.

A team from the HSE FCS AI and Digital Science Institute and Sber has developed a new CREATE architecture that combines both approaches.

Elfat Sabitov

'We figured out how to combine the strengths of transformers and graph-based models. We trained the model in two stages. First, we trained the graph component separately to analyse the connections between products. Then we added the second component, which captures the sequence of purchases, and trained the two components together, configuring them to complement each other. As a result, the model learned to use both global connections and the chronology of user actions as parts of a single mechanism,' explains Elfat Sabitov, Research Assistant at the International Laboratory of Stochastic Algorithms and High-Dimensional Inference at HSE FCS.

The researchers tested the system on five open datasets, including Amazon Reviews (233 million ratings), MovieLens (1 million), and Yambda (50 million). CREATE consistently outperformed sequential and graph-based models, as well as state-of-the-art hybrid approaches.

In tests where the model relied solely on users’ behavioural profiles, its accuracy in predicting the next product increased by an average of 12.5%. When information about product properties, categories, and relationships was added to the profiles, the recommendations became more accurate across all three evaluation criteria. The ranking metric, which measures the quality of the top recommendations, improved by 4%. Completeness increased by 5.8%, meaning that the system was less likely to miss relevant products. Coverage increased by 9.3%, making the recommendations more diverse. Thus, integrating a knowledge graph not only improves accuracy but also makes recommendations more varied and richer, qualities that end users value.

Ruslan Israfilov

'There are several areas within Sber’s ecosystem where the quality of recommendations depends not only on users’ interaction histories but also on knowledge about products, services, and their use cases. Classical recommendation models are good at identifying connections based on clicks, views, and purchases, but some important relationships cannot be reliably inferred from behavioural data alone. Knowledge graphs make it possible to explicitly represent these relationships and enrich user signals with additional context. Therefore, research demonstrating how to effectively combine graph representations with sequential models and achieve measurable improvements in the quality and diversity of recommendations is particularly valuable to us,' says Ruslan Israfilov, Executive Director of Data Research at Sber’s B2C Data and Recommendation Systems.