 |
Computational Intelligence - Learning with Neural Methods on Structured Data |
|
Computational Intelligence - Learning with Neural Methods on Structured Data |
Main contributors:
Barbara Hammer, Andreas Rechtien, Marc Strickert, Thomas Villmann (University of Leipzig)
Publications:
See publications on Barbara's or Marc's page. A download will be available soon.
Main idea:
SRNG or GRLVQ networks (see here) provide efficient and robust prototype based clustering with additional relevance terms for the input dimensions of data. If data are low dimensional or only few relevance terms are nonvanishing, this yields a visualization and hence an intuitive understanding of the underlying classifier directly. However, the situation changes if more than two or three dimensions still contribute to the classification as is often the case in practical applications.
The BB-Tree algorithms allows to extract decision trees of the classifier with little extra cost which are in practice comparable to the original classifier and hence allow insight into the underlying behavior. The BB-Tree algorithm recursively selects dimensions according to their respective relevance and splits in the dimension according to the prototypes. Rules comparable to results in the literature have been obtained in this way on several benchmarks.
Further extensions include: