Information Technology Reference
In-Depth Information
Ta b l e 2 .
Results using naıve Bayes with the
Hyper
problem
Hyper
a
/
Hyper
b
avg std dev conf
ASE
92.74 / 93.93 1.92 / 2.88 0.38 / 0.49
SE
0
.
1
92.72 / 93.92 2.20 / 2.52 0.43 / 0.50
SE
0
.
25
92.70 / 93.91 2.22 / 2.53 0.43 / 0.50
Fix
64
91.82 / 92.35 2.54 / 2.98 0.50 / 0.58
SEA
64
92.97 / 94.44 1.80 / 1.64 0.50 / 0.32
DWM
64
91.82 / 92.76 2.12 / 2.74 0.42 / 0.54
Oza
64
92.39 / 93.73 2.30 / 0.40 0.45 / 0.08
Single
90.04 / 92.68 3.25 / 2.82 0.64 / 0.55
problems, demonstrate that these kinds of approaches guarantee appreciable results only
with a quite stable phenomenon. They do not provide a fast reaction to concept drift,
since the number of models involved in the classification task is constant in time, and
when a drift occurs, they have to change a large part of the models, before classifying
new concepts correctly.
Ta b l e 3 .
Overall results with the
cHyper
problem
cHyper
a
/
cHyper
b
/
cHyper
c
- decision tree
avg std dev conf
ASE
83.58 / 88.72 / 93.19 0.51 / 0.40 / 0.28 0.10 / 0.08 / 0.06
SE
0
.
1
84.05 / 89.43 / 93.09 0.49 / 0.40 / 0.32 0.10 / 0.08 / 0.06
SE
0
.
25
78.42 / 86.10 / 91.86 0.86 / 0.35 / 0.23 0.17 / 0.07 / 0.23
Fix
64
70.26 / 82.02 / 90.62 2.58 / 1.23 / 0.13 0.51 / 0.24 / 0.13
SEA
64
70.26 / 82.14 / 90.04 2.58 / 1.10 / 0.14 0.51 / 0.22 / 0.14
DWM
64
77.75 / 85.18 / 92.65 1.94 / 0.60 / 0.14 0.38 / 0.04 / 0.14
Oza
64
81.99 / 89.60 / 92.40 0.97 / 0.37 / 0.25 0.19 / 0.07 / 0.25
Single
81.50 / 87.85 / 89.99 1.60 / 0.70 / 0.34 0.31 / 0.14 / 0.34
cHyper
a
/
cHyper
b
/
cHyper
c
-naıve Bayes
avg std dev conf
ASE
87.52 / 92.23 / 95.94 0.38 / 0.43 / 0.33 0.09 / 0.08 / 0.06
SE
0
.
1
87.62 / 92.62 / 95.98 0.42 / 0.43 / 0.47 0.08 / 0.09 / 0.09
SE
0
.
25
79.90 / 86.80 / 92.14 0.83 / 0.40 / 0.22 0.16 / 0.08 / 0.22
Fix
64
73.72 / 83.69 / 94.16 2.60 / 1.35 / 0.40 0.51 / 0.26 / 0.40
SEA
64
73.72 / 84.23 / 94.78 2.60 / 1.27 / 0.31 0.51 / 0.25 / 0.31
DWM
64
85.93 / 92.18 / 95.63 1.76 / 0.18 / 0.38 0.35 / 0.04 / 0.38
Oza
64
80.01 / 87.31 / 89.78 1.23 / 0.54 / 0.56 0.24 / 0.11 / 0.56
Single
81.25 / 89.47 / 93.34 2.02 / 0.87 / 0.84 0.40 / 0.17 / 0.84
Results with Evolving Data Sets.
Table 3 reports the overall results obtained ana-
lyzing the
cHyper
problem, considering both decision tree and naıve Bayes models.
Differently from the results obtained with stable data sets, the active model threshold
influences the overall results. Varying the value from 0.1 to 0.25, and especially con-
sidering
cHyper
a
and
cHyper
b
,
SE
system presents a difference even larger than 6%
between the two values. On the contrary, our
ASE
approach provides an accuracy in line
with the best one, even considering standard deviation. This demostrates that, without
knowing the ideal threshold value for model activation, our
ASE
approach represents
the right solution to the different situations involved in a stream scenario, and simulated