| Description: | WEASEL is composed of a supervised symbolic time series representation for discriminative word generation and the BOP model for building a discriminative feature vector. First, WEASEL extracts z-normalized windows of varying lengths from a time series. Next, each window is approximated using the Fourier transform, and only those real and imaginary Fourier values are kept that best separate time series from different classes using an ANOVA F-test (as opposed to BOSS that uses the first Fourier values). These Fourier values are discretized into a word based on a supervised symbolic transformation, which chooses discretization boundaries that best separate the classes. All window-lengths (there are in total O(n) window lengths), unigrams and bigrams are enumerated in a single feature vector (BOP), and irrelevant features are removed using the Chi-squared test. Finally, a linear time logistic regression classifier is applied that returns class probabilities. |