qxSAX - An Adaptable Symbolic Aggregate Approximation for Class Balancing and Seamless Classification
Eduardo de Medeiros da Silveira, Joaquim Assunção · 2024
Time series are broadly present in science and industry. In specific scenarios, it is useful to classify series in order to gain knowledge regarding a specific range of values. In such cases, we often use symbolic representation, as it can reduce the data dimensionality creating representative symbols, making the data discrete and allowing specialized algorithms to be applied to the data. One of the most prominent methods of this type of representation is the Symbolic Aggregate approXimation (SAX), which, in addition to generating the symbolic sequence, also reduces the data dimension. However, one of the problems of SAX is that, in order to guarantee the balance of symbols, it assumes the normality of the distribution, which fails in some distributions and causes the class imbalance problem. Furthermore, all SAX-like methods have a rigid approach to generating a symbolic output, which can be frustrating when the accuracy of the interval is desired. We propose a unique seamless approach to guarantee the balance among the classes and allow an interval selection for a desired symbol, which may lead to better results in classifying time series.