Russian scientists have presented an innovative approach to optimizing the training of artificial intelligence (AI) systems, based on classical machine learning algorithms, which will significantly improve the accuracy of recommendations in online services and marketplaces. According to the press service of Sber, the methodology solves the problem of excessive or insufficient number of training steps in modern models, which often reduces their accuracy.
The study, conducted by Bulat Ibragimov, a researcher at the AIRI Institute of Artificial Intelligence, and Gleb Gusev, director of the Sber AI Lab, showed that an adaptive approach to training gradient descent algorithms can improve the performance of recommendation systems. The researchers suggested that current training methods do not take into account the complexity and structure of different subsets of data, which reduces the quality of recommendations.
To overcome these limitations, Russian specialists have developed an algorithm that divides the data into blocks according to the level of complexity and finds the optimal training parameters for each block. This approach was tested on two popular algorithms — LightGBM and CatBoost, which are widely used in recommendation systems.
The test results showed that the adaptive training method increases the accuracy of recommendations by 2% compared to traditional methods, while slightly increasing the time and resource costs for the training process. This approach can significantly improve the quality of recommendations in various online services and marketplaces, the researchers note.
Read materials on the topic:
Artificial intelligence taught to select real estate for Russians
