Database Reference
In-Depth Information
Table 5. Average detection rate of the 24 trials (all signif-
icant attributes included in XP).
Type of change:
A
T R
Average
Number of changes:
8
4
12
24
CD
67%
99%
100%
90%
XP(candidate)
100%
100%
XP(target)
100% 100%
100%
XP(target+candidate)
99%
98%
98%
99%
Table 6. Distribution of 5% significance out-
comes in the 24 trials.
Type of change:
ATR
sum
Number of changes:
8
4
12
24
CD
0
3
12
15
XP(candidate)
4
4
XP(target)
1
9
10
XP(target+candidate)
4
3
3
10
3.3.
Results — Part 2 (Time Series Data)
Table 7 and Figure 1 describe the outcomes of applying the IFN algorithm
to seven consecutive periods from the same artificially generated database
(Database#1), with two changes introduced, and using the change detection
methodology to detect significant changes that have occurred during these
periods.
The results of the experiment can be summarized as follows:
•
All changes are detected by the change detection procedure only in the
relevant period where the corresponding change has occurred.
•
Whenever CD attained a significant level (above 5%), XP also reached a
significant value (more than 5%).
•
The effect of the change decreases after one period for CD (when the
change occurs in period
K
, the significance level in period
K
+1 is most
similar to period
1). Hence, after one period, if a change is not
properly detected, it will be absorbed and discarded.
K −
•
The XP estimator of distribution change in input and target variables is
quite sensitive to the effect of any change.
These results support the assumptions that a change in the target vari-
able does not necessarily affect the classification “rules” of the database and
that a change can mainly be detected in the first period after its occurrence.
