 |
|
Spotlight
Looking for a public data mining software? Try Orange, a scriptable component based
framework. |
| FRI > Biolab > Function Decomposition > Constructive Induction | |
Constructive InductionFunction decomposition and HINT, in a way, perform what is known as "constructive induction". From the raw data set, HINT constructs a hierarchy of new attributes and their definition. Constructive induction is therefore the very means how HINT builds the classification model from data. In the terminology of Michalski and Wnek, HINT's attribute construction may be termed as data-driven construction induction. Note, however, that most existing constructive induction tools rely on existence of constructive operators (like AND, OR, ...) and are often limited to discovery of two-valued constructs only. HINT's advantage is that it does not need constructive operators: HINT not only finds the attributes that may define good concepts, but also finds their definition. HINT can therefore be seen as data-driven operator-free constructive inducer. There are two ways HINT can be used for constructive induction. In HINT-only mode, constructive induction is the basis for the HINT's discovery of hierarchy of concepts. Alternatively, HINT may be used as preprocessor that constructs attributes for some other machine learning algorithm. In this mode, HINT may either offer any attribute from discovered concept hierarchy, or attributes that depend only on original attributes. These ideas were first explored in[2] Zupan B, Bohanec M, Demsar J, Bratko I: Feature transformation by function decomposition . IEEE Intelligent Systems & Their Applications, 1998, vol. 13, pages 38-43. where we show that attributes constructed by HINT can significantly improve performance of classification tree inducer C4.5. |
|
 |
|