A parsimonious mixture of Gaussian trees model for oversampling in imbalanced and multimodal time-series classification.

摘要: We propose a novel framework of using parsimonious statistical model, known as mixture Gaussian trees, for modeling the possibly multimodal minority class to solve problem imbalanced time-series classification. By exploiting fact that close-by time points are highly correlated due smoothness time-series, our model significantly reduces number covariance parameters be estimated from O(d(2)) O(Ld), where L is components and d dimensionality. Thus, particularly effective high-dimensional with limited instances in positive class. In addition, computational complexity learning only order O(Ln+d(2)) n+ positively labeled samples. conduct extensive classification experiments based on several well-known data sets (both single- multimodal) by first randomly generating synthetic learned correct imbalance. then compare results state-of-the-art oversampling techniques demonstrate when proposed used oversampling, same support vector machines classifier achieves much better accuracy across range sets. fact, method best average performance 30 times out 36 according F-value metric. Our also competitive compared nonoversampling-based classifiers dealing