huyp777 / efficient-shaplet-selection-for-tsc

Efficient Shaplet Selection for TSC

Time series classification (TSC)

TSC has been attracting great interest over the past decade. While dozens of techniques have been introduced, recent empirical evidence has strongly suggested that shapelets based TSC algorithms outperform many previous TSC algorithms in terms of accuracy,efficiency and interpretability. According to the concept of shapelets Shapelets are not only subsequences extracted from one time series, but also have distinctly representative characteristics of class membership. With the help of shapelets, TSC can utilize the similarity between two shapelets, rather than the similarity between two entire time series, to complete time series classification. In consequence, the overall performance of these shapelet based TSC methods can be greatly enhanced, moreover the appropriate shapelets can provide enough information to make the results of classification more explainable. Therefore, after that, an evolutionary algorithm by utilizing shapelets for TSC have been proposed.

Shapelet Transformation(ST)

ST, which not only optimizes the process of shapelets evaluation, but also allows various classification strategies(SVM,Random forest,etc.) to be adopted to classify time series objects after the shapelets selection process has been completed.

Efficient Shapelet Selection for ST

The team tried to use various acceleration strategies to accelerate the corresponding shapelet selection for traditional ST, thereby improving the overall efficiency of ST in TSC. the corresponding speedup strategies for Shaplets selection include "Finding the central sequence","Subclassing & Sample time series " , "Refining shapelet Candidates by IDPs" and etc.

Part of comparision experimental results

Comparision experiments from originial ST, ST-R(Refine strategy with IDP ) , ST-S(Selection strategy) and the ESS (utilizes all the strategies together to improve the efficiency of ST)