shapeletbased

Shapeletbased refers to a family of methods in time series analysis that rely on discriminative subsequences, called shapelets, to classify or analyze sequences. The central idea is that certain short segments within a time series carry strong information about its class, and by identifying these shapelets, one can transform raw sequences into a feature space that makes classification easier and more interpretable.

In shapeletbased approaches, the discovery process searches for subsequences of a given length that best separate classes. Evaluation criteria often measure how well a candidate shapelet distinguishes between classes, using metrics such as information gain, impurity reduction, or classification accuracy on a validation set. Due to the large search space, practical methods employ pruning, sampling, or heuristic search strategies to reduce computation. Once a collection of shapelets is found, each time series is transformed by computing the distance from the series to each shapelet, typically using the minimum distance between the shapelet and any subsequence of equal length within the series. The resulting distance-based features are then fed into standard classifiers such as k-nearest neighbors, logistic regression, support vector machines, or ensemble methods.

Shapeletbased methods offer interpretability, as selected shapelets can be inspected as concrete, human-readable patterns associated with particular classes. They have been applied primarily to univariate time series, with extensions to multivariate data and online or incremental discovery. Limitations include significant computational cost for large datasets and long sequences, as well as sensitivity to normalization and shapelet length choices. Despite these challenges, shapeletbased techniques remain influential in time series classification and related domains.