Handling Imbalanced Time Series Through Ensemble of Classifiers: A Multi-class Approach for Solar Flare Forecasting

Abstract

Solar Flares (SF) are sudden releases of large amounts of energy from the solar atmosphere [1]. They are categorized into 5 classes, namely, A, B, C, M and X, respectively in order of their strength, where SF of class A are the least harmful, while X flares are the most powerful and dangerous ones. They are categorized according to the level of X-ray emitted by Sun during the event. These phenomena impact satellite communications [2], Global Positioning System (GPS) and may also produce electricity power blackouts. So, it is imperative to develop robust solar flare forecasting systems. The features that influence the forecasting process are not known. We can find papers that use features derived from magnetogram vector [3-5], sunspot area [6], radio flux or Xray flux [7] and [8] the X-rays time series. Solar Flare datasets are extremely imbalanced. Most work in literature using traditional classifiers that deal with imbalanced datasets have the drawback of producing biased results [9-11]. An alternative to handle the poor results of the learning in imbalanced data is the usage of an Ensemble of Classifiers (EC) [12-14]. The EC main goal is to improve "weak" classification methods by applying many of weak classifiers (also called base inducers), so that the final classification may produce more accurate results. In this sense,

Solar flare forecasting

“…AdaBoost Lan et al (2012) Bayesian statistics Díscola Jr. et al (2018a,b), X. Zhang, J. Liu, and Q. , Yu, Huang, H. Wang, Cui, et al (2010) and Wheatland (2005) DecisionTrees Díscola Jr. et al (2019Jr.…”

“…et al (2019Jr. et al ( , 2018a, Huang and H.-N. Wang (2013), X. Zhang, J. Liu, and Q. , Huang, Yu, et al (2010), Yu, Huang, Q. Hu, et al (2010) and Yu, Huang, H. Wang, and Cui (2009) Ensembles Lim et al (2019), Díscola Jr. et al (2019, Guerra, Pulkkinen, and Uritsky (2015) and Huang, Yu, et al (2010) Expert systems McIntosh (1990) and Miller (1988) Extremely randomized trees (ERT)…”

“…Al-Ghraibah, Boucheron, and McAteer (2015) The least absolute Support-vector machines (SVM) Domijan, Bloomfield, and Pitié (2019), Alipour, Mohammadi, and Safari (2019), Díscola Jr. et al (2019, 2018a, Florios et al (2018), Inceoglu et al (2018), Sadykov and Kosovichev (2017), Nishizuka, Sugiura, Kubo, Den, Watari, et al (2017), Raboonik et al (2016), Bobra and Couvidat (2015), Muranushi et al (2015), Yang et al (2013), Yuan et al (2010), R. Li, Cui, et al (2008) and Qahwaji and Colak (2007) Support-vector machine regressor (SVR) Boucheron, Al-Ghraibah, and McAteer (2015) Unsupervised fuzzy clustering Benvenuto et al (2018) Unsupervised learning vector quantization R. Li, H. Wang, et al (2011) Usually, the approaches mentioned earlier aimed to improve some reference performance score, find new promising features associated with flare occurrence or increase how far in advance solar flares could be forecast. Despite most times succeeding in achieving those goals, most of them share a common aspect: using statistical or machine learning techniques to design their classifiers (CAMPOREALE, 2019).…”