Page 152 -

4.7 Statistical Classifiers in Data Mining 139

- The ill-posed nature of the data mining problems, usually with many available
solutions but no clear means of judging their relative quality This is due to: the
curse of dimensionality already mentioned in section 2.6, which is definitely a
relevant phenomenon in many data mining applications; the incorrect inferences
from correlation to causality; the difficulty in measuring the usefulness of the
inferred relations.
The most important statistical classifier approach in data mining is the decision
tree approach. As a matter of fact, tree classification can be achieved in a very
efficient way, e.g., by using the previously described CART approach, and can also
provide important semantic information, especially when using univariate splits.
The contribution of individual features to the target classification, which is a top
requirement in data mining applications, is then easily grasped.
In order to get a taste of the application of tree classification to a typical data
mining problem, and to discuss some of the issues that have just been enumerated,
we will consider the problem of assessing foetal heart rate (FHR) variability
indexes in the diagnostic of a pathological foetal state responsible for a "flat-
sinusoidal" (FS) tracing. For this purpose, we will use the CTG dataset with 2126
cases (see Appendix A). Four variability indexes are used (MLTV, MSTV, ALTV,
ASTV), which measure the average short-term (beat to beat) and average long-
term (in 1 minute sliding window) variability of the heart rate, as well as the
percentage of time they are abnormal (i.e., below a certain threshold).
Using the CART design approach available in Statistics, we obtain the decision
tree of Figure 4.45, which uses only two variability indexes (percentage of time the
short-term and long-term variabilities are abnormal).

Figure 4.45. Decision tree for the flat-sinusoidal class (FS) in 2126 FHR tracings.
FS cases correspond to the dotted rectangles, non-FS to the solid ones.