MIL OSI Translation. Region: Russian Federation –
Source: State University Higher School of Economics – State University “Higher School of Economics” – Source: State University “Higher School of Economics” – Source: State University “Higher School of Economics” – State University “Higher School of Economics
Researchers fromof the Center for Artificial Intelligence иcomputer science faculty National Research University Higher School of Economics presented a new algorithm for detecting structural changes in time series. The method uses a neural network to compare different segments of a series, which allows faster detection of changes in its behavior. The results of the work weresubmissions At the 26th International Conference on Artificial Intelligence and Statistics – AISTATS (A*).
The study was supported by a grant for research centers in the field of artificial intelligence provided by the Analytical Center under the Government of the Russian Federation.
In modern machine learning tasks, it is often necessary to process time series, i.e. sequences of time-ordered observations. In this case, the data can be of different nature: from the number of COVID-19 strain patients and monitoring indicators of patients undergoing rehabilitation after stroke, to the hourly number of posts in social networks on a particular topic and seismic activity sensor readings.
The frequency with which new data come in such observations can vary considerably. But there is a common feature: sudden changes in the behavior of these time series can signal an important event – the beginning of a new wave of a pandemic, the need to provide urgent care to a patient, an earthquake, and so on. Their timely detection will help prevent or at least mitigate undesirable consequences.
The moment in time when the data no longer fit the expected pattern or trend is called a discordance. It is worth noting that important structural changes in the sequence of observations are not always noticeable to humans. This leads to the need to develop automatic methods for their detection.
The problem of discrepancy detection has long been one of the classical problems in mathematical statistics, so researchers all over the world are working on creating effective methods for analyzing data and detecting structural changes. One of such methods, an algorithm for detecting discrepancies in time series, was developed by researchers at the Computer Science Department of the National Research University Higher School of Economics (HSE)Nikita Puchkin and Valeria Scherbakov.
There are several ways to detect discordance in time series, and they can be categorized into groups depending on what kind of structural change is to be detected. Some methods focus on changes in mean values, others on changes in trend or on the volatility of the data (a measure of how much the data change over time). There are also methods that can detect discordance of arbitrary kinds, i.e. non-parametric methods. These are particularly useful when the effects of an event have not yet fully manifested themselves, the trend and volatility of the time series remain the same, but there are changes in other characteristics of the data. Understanding these methods helps researchers and analysts to more accurately identify discordance in time series and take appropriate action.
Scientists note that in a number of studies, nonparametric methods of fault detection are presented without theoretical estimates of the speed of detecting changes in the sequence of observations, which raises questions about the reliability of the results. Therefore, the researchers of the Center for Artificial Intelligence of the National Research University Higher School of Economics set an ambitious task to develop a method that, on the one hand, would be practical and, on the other hand, would have a clear theoretical justification.
Our algorithm is based on a simple idea: if the behavior of the time series has changed, the observations before and after the moment of discord can be distinguished from each other. For this purpose, we use a neural network, optimizing its weights in such a way that the contrast between the parts of the sample before and after the debugging is the most pronounced. Therefore, the method is universal, and most importantly, its efficiency is confirmed mathematically.
To verify the algorithm’s performance, the scientists conducted a series of tests of varying complexity, comparing it to several popular nonparametric fault detection methods. The tests took into account how often the algorithm makes errors, producing false signals, and how long it takes to detect changes. As a result, the algorithm showed promising results, detecting important events or changes in data on average 30% faster than competitors.IQ
Subscribe to IQ.HSE
Please note; This information is raw content directly from the source of the information. It is accurate to what the source states and does not reflect the position of MIL-OSI or its customers.
Please note; This information is raw content directly from the information source. It is accurate to what the source is stating and does not reflect the position of MIL-OSI or its clients.
EDITOR’S NOTE: This article is a translation. Apologies should the grammar and or sentence structure not be perfect.