Investigating feature extraction techniques for imbalanced time-series data

Abstract

High class data imbalance is usually present in many applicatio, such as fraud detection and cancer diagnosis, hence effective classification with time - series data is an essential topic of study. Furthermore, excessively imbalanced data presents a challenge, since most learners will be biased toward the majority group, and in extreme circumstances, will overlook the minority group completely. Over the previous two decades, fundamental methodologies have been used to study class imbalance in depth. Despite recent breakthroughs in addressing data imbalance with feature extraction and its growing popularity, there is relatively little empirical work in the domain of feature extraction with time-series based class imbalance. Following record-breaking performance outcomes in various complicated domains, researchers are now looking into the usage of feature extraction approaches for issues with significant degrees of class imbalance. In order to better understand the effectiveness of feature extraction when applied to class imbalanced data, available research on class imbalance, feature extraction, and fundamental approaches like SMOTE, Resampling, and others are examined. This study explores into the specifics of each study's execution and experimental outcomes, as well as providing more insight into its advantages and limitations. We discovered that there is relatively limited study in this field. Several classic approaches for class imbalance, such as data sampling and SMOTE, work with feature extraction, but more sophisticated methods that take use of minority class feature learning abilities potential application. The survey continues with a discussion that identifies numerous gaps in time-series data based on class imbalanced data in order to improve future study.