Machine Learning in Classification Time Series with Fractal Properties

Lyudmyla Kirichenko,Vitalii Bulakh,Tamara Radivilova

doi:10.3390/data4010005

Lyudmyla Kirichenko, Vitalii Bulakh + Show 1 more

Open Access

https://doi.org/10.3390/data4010005

Copy DOI

Journal: Data	Publication Date: Dec 28, 2018
Citations: 40	License type: CC BY 4.0

Affiliation: Kharkiv National University of Radio Electronics

Abstract

The article presents a novel method of fractal time series classification by meta-algorithms based on decision trees. The classification objects are fractal time series. For modeling, binomial stochastic cascade processes are chosen. Each class that was singled out unites model time series with the same fractal properties. Numerical experiments demonstrate that the best results are obtained by the random forest method with regression trees. A comparative analysis of the classification approaches, based on the random forest method, and traditional estimation of self-similarity degree are performed. The results show the advantage of machine learning methods over traditional time series evaluation. The results were used for detecting denial-of-service (DDoS) attacks and demonstrated a high probability of detection.

Highlights

Over the past two decades, machine learning methods for time series have been proposed and developed, which are used for many tasks of time series analysis, including classification [1,2,3,4]
The classification was performed for time series with different multifractal properties
Training time series for the experiment were obtained by the generation of multifractal cascades with weights obtained by symmetric beta distribution

Summary

Introduction

Over the past two decades, machine learning methods for time series have been proposed and developed, which are used for many tasks of time series analysis, including classification [1,2,3,4]. Many complex technical and information systems have a fractal (self-similar) structure, and their dynamics is represented by time series with fractal properties [5]. For such systems, there are problems of recognition and classification of fractal series. There are problems of recognition and classification of fractal series Most often, they are solved by evaluation and analysis of self-similar properties. Another paper [8] reported a novel method on the machine learning based classification of fractal features of time series using twin support vector machines. The authors of [9] demonstrated the successful application of the

Objectives

Methods

Results

Conclusion