Method trees: building blocks for self-organizable representations of value series: how to evolve representations for classifying audio data
share
In this paper we introduce a framework for automatic fea- ture extraction from very large series. The extracted fea- tures build a new representation which is better suitable for a given learning task. The development of appropriate feature extraction methods is a tedious eort, particularly because every new classication task requires tailoring the feature set anew. Therefore, the simple building blocks de- ned in our framework can be combined to complex feature extraction methods. We employ a genetic programming ap- proach guided by the performance of the learning classier using the new representation. Our approach to evolve rep- resentations from series data requires a balance between the completeness of the methods on one side and the tractabil- ity of searching for appropriate methods on the other side. In this paper, some theoretical considerations illustrate the trade-o. After the feature extraction, a second process learns a classier from the transformed data. The practi- cal use of the methods is shown by two types of experiments in the domain of music data classication: classication of genres and classication according to user preferences.
READ FULL TEXT
