О воспроизводимости анализа временных рядов на основе матричных профилей

Программирование. 2026. № 4. С. 50–66.

Брычков М. Е., Незнанов А.А.

The matrix profile (MP) quickly became one of the most important time series preprocessing methods when it was introduced in 2016, facilitating the solution of a wide range of time series analysis problems, in particular anomaly and pattern detection problems. The high significance has led to the emergence of various tools for MP calculating, but it turned out that the implementation features of the methods affect the results, which cease to be identical. The question arises as to how significant these differences are for a researcher in a laboratory and an engineer in industry when implementing time series analysis methods using MP. As shown by large-scale computational experiments, popular methods for calculating MP and their software implementations have characteristics that lead to different results for identical input data. Some differences affect only the work of researchers, while others also affect the work of data engineers and analysts. For an adequate understanding of the consequences, the work provides illustrative examples of calculating MP, a comparison of methods, and recommendations.

Keywords: временной ряд, вычислительный эксперимент, time series, воспроизводимость, computational experiment, reproducibility, matrix profile, матричный профиль