Shapelets are discriminative subsequences of time-series that best predict the target class value. ^[Guidotti et al. - Explaining Any Time Series Classifier]
A shapelet of length is an ordered sequence of values, where is the length of the original sequence. is the set of the -most informative shapelets, where the number of shapelets in the set is obtained using an heuristic.
In the paper by Grabocka et al. ^[Grabocka et al. - Learning Time-Series Shapelets.pdf] a method to learn the optimal set of shapelets without exploring all possible candidates is proposed. The procedure starts with rough initial guesses for the shapelets in S, and then iteratively optimizes the shapelets by minimizing a classification loss function using stochastic gradient descent.
tags: ai-explainability