Hier finden Sie wissenschaftliche Publikationen aus den Fraunhofer-Instituten.

Vector quantization based learning algorithms for mixed data types and their application in cognitive support systems for biomedical research

: Zühlke, Dietlind

Volltext urn:nbn:de:0011-n-2173388 (34 MByte PDF)
MD5 Fingerprint: 7f5308187c898060ba0f26a0362cc9fe
Erstellt am: 16.11.2012

Groningen: Atto Producties Europe, 2012, XXI, 253 S.
Zugl.: Groningen, Univ., Diss., 2012
ISBN: 978-90-9027048-7
Dissertation, Elektronische Publikation
Fraunhofer FIT ()
machine learning; life science informatics; learning vector quantization; mixed data types; cognitive support system; biomedical research

Multi-layer models are of increasing importance in biomedical research. By representing objects as ensembles of heterogeneous partial aspects they allow modeling complex relations. Due to this high complexity the aid of computers is needed in investigating these relations. As often there are no clear hypotheses on the expected relations, traditional bio-statistical approaches are unsuitable for this task. This thesis introduces a framework that optimizes the intrinsic grouping (clustering) of multi-layer objects as well as the grouping of these objects according to a set of given class assignments (classification). It identifies prototypical representatives of the groups. Adequate corresponding distance measures take account for the heterogeneity of the partial aspects. Additionally, the framework allows analyzing the relevance of single aspects in the models in their joined context. In the integrative analysis either preselected aspects or the whole ensemble are used for a suitable grouping. Thereby single aspects are weighted according to their influence on the grouping. These weights can under certain constraints be interpreted as relevance values. Applying the framework for heterogeneous data in breast cancer research during the thesis it could be shown that if handled suitably it succeeds as a cognitive support system for biomedical research. The identification of prototypes as well as the determination of relevance values relates to cognitive models of human expert thinking. They can be handled intuitively by the domain experts. The support system thus enables the g eneration of hypotheses concerning biomedical relations that are then testable using traditional bio-statistical approaches.