Information Technology Reference
In-Depth Information
Minority class
cluster 2
S
5
S
1
S
1
S
2
S
5
S
2
S
4
Minority class
cluster 1
S
4
Δ
1
S
3
S
3
Δ
2
Δ
(a)
(b)
S
1
S
5
S
2
S
4
S
3
(c)
Figure 7.1
The selective accommodation mechanism: circles denote current minority
class examples and stars represent previous minority class examples. (a) Estimating sim-
ilarities based on distance calculation, (b) potential dilemma by estimating distance, and
(c) using the number of current minority class cases within
k
-nearest neighbors of each
previous minority example to estimate similarities.
knowledge of the class concept. However, sole maintenance of a hypothesis or
hypotheses on the current training data chunk is more or less equal to discarding
a significant part of previous knowledge, as knowledge of previous data chunks
can never be accessed again either explicitly or implicitly once they have been
processed.
One solution for addressing this issue is to maintain all hypotheses built on
training data chunks over time and apply all of them to make predictions on
datasets under evaluation. Concerning the way in which hypotheses combine,
Gao et al. [31] employed a uniform voting mechanism to combine the hypotheses
maintained this way as it is claimed that in practice the class concepts of datasets
under evaluation may not necessarily evolve consistently with the streaming train-
ing data chunks. Putting aside this debatable subject, most work still assumes that
the class distribution of the datasets under evaluation remains tuned to the evolu-
tion of the training data chunks. Wang et al. [32] weighed hypotheses according
to their classification accuracy on
current training data chunk
. The weighted
Search WWH ::
Custom Search