The classification of high dimensional data like time series often requires the efficient extraction of meaningful features. This plugin provides virtually all methods which are known today for the feature extraction from value series. A special highlight is the completely automatic feature extraction by means of genetic programming.

Features

The methods of value series analysis are divided into groups and systematized. Together with an extended concept of windowing these methods build method trees for feature extraction, just like the operator trees already known from RapidMiner (formerly YALE). It can be theoretically shown that the windowing of efficient methods also is efficient, the same applies for method trees. Therefore we use these method trees for fast and efficient feature extraction from large value series.

The systematization and the efficiency of the methods allow automatic approaches to extract an optimal set of features from value series. This plugin includes an approach based on genetic programming. The individuals are method trees which work on a subset of the data and are mutated and recombined. The result is a operator tree which can be used on the complete data set for feature extraction.

The transformation from raw series data to feature vectors was successfully applied to different tasks in the domain of audio data, which can be seen as time series with an extraordinary length.

The theoretical foundation of this plugin and some of its methods are described in the following article: Mierswa, Ingo and Morik, Katharina. Automatic Feature Extraction for Classifying Audio Data. In Machine Learning Journal, Vol. 58, pages 127-149, 2005.

Download and Documentation

The following files are available from the RapidMiner Plugins download page: