Abstract: |
In this paper we introduce a framework for automatic featureextraction from very large series. The extracted features build a newrepresentation which is better suitable for a given learning task. Thedevelopment of appropriate feature extraction methods is a tediouseffort, particularly because every new classification task requirestailoring the feature set anew. Therefore, the simple building blocksdefined in our framework can be combined to complex feature extractionmethods. We employ a genetic programming approach guided by the performance of the learning classifier using the newrepresentation. Our approach to evolve representations from seriesdata requires a balance between the completeness of the methods on one side and thetractability of searching for appropriate methods on the otherside. In this paper, some theoretical considerations illustrate the trade-off. After the featureextraction, a second process learns a classifier from the transformeddata. The practical use of the methods is shown by two types ofexperiments in the domain of music data classification: classificationof genres and classification according to user preferences. |
