The End of Blind Search: HSE Researchers Learn How to Pick the Right AI for Recommendations
Researchers at the Institute for Artificial Intelligence and Digital Sciences at the HSE University Faculty of Computer Science in Moscow have developed an approach that makes it more efficient to select recommendation algorithms. Different methods compete against one another in pairwise comparisons, and the results of all the matchups are used to produce an overall ranking. This reduces the number of algorithms developers need to test when building new services, saving both money and time. The research was presented at the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD 2026).
Every day, millions of people rely on recommendation algorithms for everything from choosing a movie for the evening to buying a product on a marketplace. But the convenience the algorithms offer users conceals a huge challenge for tech companies: How can they determine which of hundreds of neural networks is best suited to their particular data? Researchers at the Institute for Artificial Intelligence and Digital Sciences at the HSE University Faculty of Computer Science have found an answer. Their approach replaces guesswork with the ability to accurately predict an algorithm’s performance before expensive testing begins.
A Tournament of Algorithms
The HSE team’s key finding is that there is no “golden ratio” in machine learning. No universal algorithm can recommend products, music and news equally well.
To demonstrate this and build a tool for making that choice, the researchers adapted the Bradley–Terry model. Originally used in sports to estimate the relative strength of competitors in pairwise contests, the model gave the HSE researchers a way to turn 14 popular recommendation algorithms into virtual “athletes” and pit them against one another in a tournament.
The results confirmed their hypothesis: The leaders change with the “rules of the game.” SASRec and GASATF, for example, win decisively on sequential data, where the order of a user’s actions matters. When there is no clear sequence, however, LightGCN and ALS came out on top.
Prediction Instead of Brute Force
The most valuable part of the research is its practical application. Companies typically have to train dozens of models, consuming enormous computing resources, and then run costly A/B tests with real users.
The HSE method offers a way to predict the outcome before going through that process. The system analyzes the characteristics of a new dataset and produces an advance ranking of the algorithms most likely to perform well. Experiments showed striking accuracy: The algorithm ranked first by the system actually finished among the top three after full testing 78% of the time. By comparison, a conventional approach based on averaging performance metrics achieved just 16% accuracy. In practice, businesses could eliminate clearly weaker options at the outset, sharply reducing costs.
Industry Demand: From Sber to VK
The HSE researchers’ work closely tracks a challenge the Russian tech industry has been grappling with in recent years. As early as 2023, experts at Yandex Education highlighted the difficulty of choosing the right methods for specific tasks. In 2024, Sber’s department at HSE made the problem explicit: A model that succeeds scientifically does not necessarily improve business metrics in Prometheus, so the industry needs ways to compare algorithms before deploying them.
The scale of the challenge is clearest in the numbers. In 2025, VK launched the VK RecSysChallenge, giving developers a dataset containing 40 billion user interactions. Processing data at that scale through blind trial and error is an unaffordable luxury. The HSE method provides a filter that makes it possible to work with big data more deliberately. That could be particularly important for e-commerce, fintech, streaming services and social networks.
A Russian Method on the Global Stage
The research’s scientific rigor and practical value have received international recognition. The work was presented at ACM SIGKDD KDD 2026, one of the world’s most prestigious conferences in data science. And this is more than another research paper. It is a ready-to-use methodology that is not tied to the specifics of the Russian market. Model selection is a universal problem. That gives the work international potential rooted in technological leadership: Developers of recommendation systems outside Russia can now draw on research produced by Russian scientists.
The HSE research marks an important shift. The field is moving away from an endless race to see who can build the more complex neural network and toward an era of rational engineering. For users, that could mean more accurate social media feeds, better selections of banking products and more relevant recommendations from online streaming platforms, which, according to the HSE Institute for Statistical Studies and Economics of Knowledge, are regularly used by 60% of Russian internet users. For businesses, it offers an opportunity to stop burning through budgets on blind testing and start building AI services faster, more cheaply and more systematically.
