Algorithms
SAX-VSM and its building blocks — z-normalization, PAA, SAX, tf·idf and cosine similarity — built up step by step.
This module builds SAX-VSM from the ground up. The pages are ordered: each algorithm below builds on the previous ones, ending with the full classifier and a bonus visualization technique.
Z-normalization of time series
Why and how time series are standardized to zero mean and unit variance before discretization.
readingPiecewise Aggregate Approximation (PAA)
Dimensionality reduction by frame averaging, and the lower-bounding PAA distance.
readingSymbolic Aggregate Approximation (SAX)
How SAX turns a real-valued time series into a string: PAA frames, Gaussian breakpoints, and the lower-bounding MINDIST — with a fully worked example.
readingTime series discretization via sliding window
How a sliding window turns one long time series into an ordered sequence of SAX words describing its local features.
readingSAX numerosity reduction
Collapsing runs of identical consecutive SAX words — smaller output, faster algorithms, and naturally variable-length patterns.
readingTF-IDF weighting for time series (tf·idf)
How term frequency and inverse document frequency turn per-class bags of SAX words into class-characteristic weight vectors.
readingCosine similarity for time series classification
The angle-based similarity measure that scores an unlabeled time series against each class's tf·idf weight vector.
readingInterpretable time series classification with SAX-VSM
The full SAX-VSM algorithm: per-class bags of SAX words, tf·idf weighting, and cosine-similarity classification.
readingSAX bitmap patterns visualization
Turning SAX word frequencies into color bitmaps for at-a-glance comparison of large time series collections.