Database Reference
In-Depth Information
Change-Detection Procedure
Stage 1:
For periods
K
1
build
M
using DM algorithm G.
K
1
D
Define
.
K
1
val
)
Count
n
D
.
K
1
K
1
val
)
e
ˆ
Calculate
according to V.
M
K
K
,
1
1
K
n
for every candidate and target
variable existing in periods (1,...,
x
Calculate
,
iK
1
1
K
1
).
Stage 2:
For period
K
, define
d
.
K
n
d
Count
.
K
K
e
ˆ
Calculate
according to V.
M
K
,
K
1
ˆ
d
2
ˆ
ˆ
ˆ
d
2
H
z
D
V
d
ABS(
e
e
)
V
,
Calculate
,
.
M
,
K
M
,
K
1
0
1
K
1
K
1
2
Calculate and return CD(Į).
Stage 3
:
For every candidate and target variable existing in periods
(1,...,
2
p
x
n
X
K
1
) and in period
K
calculate:
,
, and
.
iK
K
Calculate and return XP(Į).
It is obvious that the complexity of this procedure is at most O(
n
K
). It is very
easy to store information about the various distributions of the target and candidate
variables to simplify the change-detection methodology.
Using the outputs of the methodology the user can make a distinction among
the eight possible variations of a change in the data-mining classification model.
According to this new information the user of the new methodology can act in
several ways : For example, the user can reapply the algorithm from scratch
absorbing the new period and using the same incremental algorithm, making
1
c
KK
and performing the procedure again. The user can also investigate the
type of the change and its magnitude and effect on the other characteristics of the
DM model, and incorporate other known methods dealing with the specific
detected changes. One can also incorporate multiple-model approaches such as
weighting, arbitrage, and combining methods, and use the prior knowledge of the
change.
The methodology is not restricted to databases with a constant number of
variables. The basic assumption is that if the addition of a new variable influences
the connection between that target variable and the candidate variables in a
manner that inflicts on the validation accuracy (V is the method to select
Search WWH ::
Custom Search