Segmentation of Continuous Data Streams Based on a Change Detection Methodology - Advanced Techniques in Knowledge Discovery and Data Mining - page 107

3. A change in the “patterns” (rules) that define the relationship of the inputs to
the target variable, that is, a change in the model M. For instance, in the case of
examining the rate of failures in a seminar exam based on the characteristics of
the students in the course in consecutive years : if in 1999 male students
produced 60% of the failures and female students produced 5%, and in 2000 the
state was opposite, then it is obvious that there was a change in the patterns of
behavior of the data-mining model. This suggests a significant change in a
data-mining model M by the following definition: A change C is encountered
during period K if the validation error of the model M K -1 (the model that is
based on D K -1 ) on the database D K -1 is significantly different from the
validation error rate of the model M K -1 over d K . d K consists of the
accumulated set of records in period K .
4. Instability of the DM algorithm. The basic assumption is that the DM
algorithms the users choose to achieve their goal (their decision) were proven to
be stable and therefore are not likely to be a major cause for a significant
change in the model. Although this option should not be omitted, this work
does not intend to deal with stability or instability measures of existing DM
algorithms, rather than point out that an assumption of the stability of a chosen
algorithm should be set before implementing any decisions based on the
induced model. As indicated in the next section, the DM algorithm that was
used in our experiments (IFN) was proven to be stable.

As noted in Section 4.2.1 the data-mining classification model is generally
described as: M G , that is, finding the “right” connection between A and

T . The first cause is explained via a change in the set of attributes A and can also
cause a change in the target variable (for example, if the percentage of women
rises from 50% to 80% and women drink more white wine than men, then the
overall percentage of white wine consumption ( T ) will also rise) and can also
affect the overall error rate of the data-mining model if most “laws” were
generated for men. Also, if the cause of a change in the relevant period is a change
in the target variable (for example, France has stopped producing white wine, and
the percentage of white wine consumption has dropped from 50% to 30%), it is
possible that the laws between the relevant attributes to the target variable(s) will
be affected.
Therefore, it is not obvious to state which cause of change will be more
significant than others. Alternatively, this work states that there are a variety of
possible changes that might occur in a relevant period. This change can be caused
by one of the three causes mentioned earlier or by a simple combination of them.
Table 4.1 consists of the possible combinations of significant causes in a
relevant period.

: A T