Explainable time series classification with X-ROCKET

A three-part series of articles that I recently released discusses the state of time series classification, and introduces an extension to a state-of-the-art encoding model to add explainability.

The first part provides an overview over time series classification problems and one solution to them — the ROCKET encoder model. Part two explains more thoroughly how this model uses convolutions to filter input time series for a fixed set of patterns, which can be made explainable if carefully put together. Finally, part three of the article demonstrates the predictive performance and interpretability of the model in a practical usage example.

The code belonging to the described X-ROCKET model can be found in this repository.