Join our Newsletter
Contact
webmaster: Sven F. Crone
Centre for Forecasting
Lancaster University
Management School
Lancaster LA1 4YF
United Kingdom
Tel +44.1524.592991
Fax +44.1524.844885
eMail sven dot crone (at)
neural-forecasting dot com
Data driven methods from computational intelligence share the common approach of learning machines in classification for data mining Let all relevant and measurable attributes of an object, e.g. a customer, be combined in a vector x and the set 
denotes the input space with n objects. Each object belongs to a discrete class 
and we will refer to a pair 
as an example of our classification problem. Presuming that it is impossible to model the relationship between attribute vector x and class membership y directly, either because it is unknown, to complex or the data is corrupted by noise, and that a sufficient large set of examples 
is available, we can incorporate a machine to learn the mapping between x and y. The learning machine is actually defined by a set of possible mappings 
, where the functions 
themselves are labeled by the adjustable parameter vector 
The objective is to modify the free parameters to find a specific learning machine which captures the relationships in the training examples, 
, incrementally minimizing a given objective function and generalizing the problem structure within to allow correct estimation of unseen objects on the basis of their attribute values 
.
For most questions of parameterisation only rules of thumb are known. Answers that are valid for all kinds of problems cannot be given – each problem needs its own. Therefore, knowledge not only in the field of soft computing but also in the problem domain is necessary. Following, we outline the specific modelling-properties for classification for alternative network paradigms. For a comprehensive discussion readers are referred to