AIRI: Framework for improving the accuracy of recommendation systems

2024: Introducing a framework to improve the accuracy of recommendation systems

Scientists at the AIRI Institute of Artificial Intelligence have developed a framework that combines the two approaches used for May 2024 everywhere in recommendation systems. The company announced this on May 23, 2024.

Recommendation systems are tools based on AI technology that offer users personalized recommendations based on their history of actions, preferences, choices and a variety of other characteristics. These solutions help businesses increase conversion and sales volumes by offering products that match the individual interests of each particular customer.

Along with business, algorithms of recommendation systems are actively used by scientific organizations in the field of biological and medical chemistry, chemical-pharmaceutical and biological research, including in the selection of the most promising molecules with antiviral activity.

Researchers at the AIRI Institute have found a way to combine the two approaches most used in recommendation systems: sequence-based learning and graph-based learning. Work on creating and testing the framework took a year. The experiments were conducted on four widely used open data sources.

Algorithms based on consistent training in the formation of recommendations use information not only about products selected by users, but also about the order of their consumption. In turn, the graph approach allows you to take into account the long-range connections between goods with which the user did not directly interact. In this case, predictions are generated based on the selection of other users who have previously interacted with the same goods. The key difficulty in combining the two approaches into a single framework was the need to prevent errors in the methodology when taking into account the order of connections. This is what often leads to "leaks from the future" - features of the generalization of behavioral patterns, in which the recommendations take into account goods that the user could not see due to their later appearance.

The framework we created significantly improves the accuracy of recommendations in the datacets used by the market. We plan to continue to work on improving it so that, along with quality, efficiency increases. This approach will be useful for machine learning researchers in various fields: from medicine to the entertainment industry, - said Evgeny Frolov, PhD, head of the scientific group "Personalization Technologies" at the AIRI Institute.