Semantics-Preserving Dimensionality Reduction in Intelligent Modelling

Abstract

The last two decades have seen the exponential increase of powerful model-building methods being called upon to solve large-scale real-world problems. However, for all their accuracy, many of the systems built using these methods tend to be black boxes. Inspecting the model is difficult, and they are often unable to provide explanations about their reasoning. This limits the trust human users can put in such systems. Modelling methods that preserve their semantics in human-readable, linguistic terms are thus very desirable. This, however, is a moot point when very complex domains are involved. Even linguistic rules can become too complicated for humans to follow if the domain they model is more complex than the human mind can handle. This chapter discusses an approach for semantics-preserving dimensionality reduction, or feature selection, that simplifies domains in the context of fuzzy or neural modelling, all the while retaining the accuracy of the respective model. The approach is described with respect to a number of example applications.