Feature Extraction over Multiple Representations for Time Series Classification

We suggest a simple yet effective and parameter-free feature construction process for time series classification. Our process is decomposed in three steps: (i) we transform original data into several simple representations; (ii) on each representation, we apply a coclustering method; (iii) we use coclustering results to build new features for time series. It results in a new transactional (i.e. object-attribute oriented) data set, made of time series identifiers described by features related to the various generated representations. We show that a Selective Naive Bayes classifier on this new data set is highly competitive when compared with state-of-the-art times series classification methods while highlighting interpretable and class relevant patterns.