A novel shape-based time series classification with SAX-Ensemble
Mariem Taktak, Slim Triki · International Journal of Computer Applications in Technology · 2023
Since the first publication of the Symbolic-Aggregate Approximation (SAX), a lot of extensions with novel SAX-distance measure are published. Each of them attempts to integrate additional statistical features in order to improve original SAX average-based feature. Each SAX-feature has its own distance function which quantifies the (dis)similarity between two Time Series (TS). However, none of them can fit the overall shape-characteristics of a TS and give the superiority to an individual SAX-based classifier. In order to combine the prediction of each single SAX-based classifier, we propose a collection of several SAX-features to compose a shape-based ensemble for TS classification. The proposed SAX-Ensemble scheme is applied on a multiple domain representation of the TS where the diversity of collected SAX-features make the setting of the SAX-discretisation parameters a challenging task especially for a long TS data or a large training data set. In order to avoid a time-consuming of either grid search or expensive optimisation algorithm, we instead apply a data-aware or data-agnostic parameters setting technique. Experimental results on real TS database show that the performance of the proposed SAX-Ensemble with data-aware technique exceeded the SAX-based classifiers with more flexible and realistic parameters estimation.