Scientists Present New Solution to Imbalanced Learning Problem

© iStock

Specialists at the HSE Faculty of Computer Science and Sber AI Lab have developed a geometric oversampling technique known as Simplicial SMOTE. Tests on various datasets have shown that it significantly improves classification performance. This technique is particularly valuable in scenarios where rare cases are crucial, such as fraud detection or the diagnosis of rare diseases. The study's results are available on ArXiv.org, an open-access archive, and will be presented at the International Conference on Knowledge Discovery and Data Mining (KDD) in summer 2025 in Toronto, Canada.

The problem of imbalanced learning is becoming increasingly relevant across various fields, including banking and medicine. Conventional methods, such as random oversampling, often generate low-quality samples or fail to accurately model rare class data.

Simplicial SMOTE (Synthetic Minority Oversampling Technique), a novel solution proposed by scientists from HSE University and Sber AI Lab, addresses these issues by enabling more accurate modelling of complex topological data structures and improving classifier performance on imbalanced datasets.

It generates new examples of a rare class by leveraging information from multiple closed instances ('simplex'), rather than just two close points, as in the original SMOTE and its well-known modifications. This facilitates a better understanding of the data and advances performance. The technique improves training on imbalanced data, where one class (eg, normal transactions) has many examples, while another class (eg, fraud) has few.

Researchers have experimentally shown on a large number of test datasets that the proposed approach achieves significantly better performance metrics, such as the F1 Score and Matthews Correlation Coefficient, for both the basic SMOTE and its modifications. In particular, an improvement was observed in gradient boosting, a classifier commonly used in practice.

Andrey Savchenko

'Our technique is particularly effective for tasks involving imbalanced data, where the rare class holds greater significance. Banks can use Simplicial SMOTE to detect fraud more effectively, and medical centres can apply it to diagnose rare diseases,' says Andrey Savchenko, co-author of the article and Leading Research Fellow at the Laboratories for Theoretical Modelling in AI of the HSE AI and Digital Science Institute.

The new technique can be integrated into existing oversampling algorithms (such as Borderline-SMOTE, Safe-level-SMOTE, and ADASYN), enabling better accuracy without significantly increasing computational complexity. According to the researchers, the developed approach could contribute to the creation of more accurate and reliable machine learning models, thereby improving the quality of analytics.

The study was conducted with support from the HSE Basic Research Programme.