Database Reference
In-Depth Information
(iv) The effect of a change is relevant and can be mainly detected in
the period of change. If not detected, the influence of a change in
a database can be absorbed in successive periods.
(v) All procedures, hypothesis tests and definitions, were validated on
artificially generated databases and on a real-world database.
(vi) The change detection procedure has low computational costs. Its com-
plexity is
O
(
n
K
), where
n
K
is the total number of validation records in
periods, since it requires only testing whether the new data agrees
with the model induced from previously aggregated data.
(vii) The CD metric can also be used to determine whether an incremen-
tally built model is stable. In our real-world experiment, stability of
info-fuzzy network has been confirmed by the increasing confidence
levels over initial periods of the data stream.
K
The Change Detection Procedure, with the use of the statistical estima-
tors, can detect significant changes in classification models of data mining.
These changes can be detected independently of the data mining algorithm
used (e.g., C4.5 and ID3 by Quinlan, KS2, ITI and DMTI by Utgoff, IFN
by Last and Maimon, IDTM by Kohavi, Shen's CDL4, etc. ), or the induced
classification model (rules, decision trees, networks, etc.). As change detec-
tion is quite a new application area in the field of data mining, many future
issues could be developed, including the following:
(i) Implementing meta-learning techniques according to the cause(s) and
magnitude(s) of a change(s) detected in period
K
for combining sev-
eral models, such as: exponential smoothing, voting weights based on
the
CD
confidence level, ignoring old or problematic periods, etc.
(ii) Increasing the granularity of significant changes. There should be sev-
eral sub-types of changes with various magnitudes of the model struc-
ture and parameters that could be identified by the change detection
procedure, in order to give the user extra information about the mined
time series.
(iii) Integrating the change detection methodology in an existing data min-
ing algorithm. As indicated above, the change detection procedure's
complexity is only
O(n)
. One can implement this procedure in an
existing incremental (online) learning algorithm, which will continue
eciently rebuilding an existing model if the procedure does not indi-
cate a significant change in the newly obtained data. This option is
also applicable for meta-learning and multi-model methods.
