Information Technology Reference
In-Depth Information
Table 3.
Relative Absolute Errors for the Homogeneous environment
Strategy
RandomR
SVM-RFE
GELF
M5P
34.7%
22.0%
31.2%
ANN
34.1%
20.8%
41.9%
k-NN
35.1%
21.2%
32.0%
SVR
43.1%
24.8%
40.9%
Ensemble model
36.0%
21.1%
20.7%
(that had an average performance for the rankers) achieved the lowest error
for GELF. This behavior evidences an important characteristic of the ensemble
method, i.e. its robustness. The method can be applied to different scenarios
resulting always in a good performance.
Heterogeneous Environment.
Table 4 shows the errors for the heteroge-
neous environment. Highlighted values represent the minimum errors for each
type of task. For the RandomR task, the best performing methods are k-NN and
Bagging-M5P with a 19.8% error. The predictions for this (heterogeneous) envi-
ronment are much more accurate than in the previous case. For the SVM-RFE
task, the best performance is achieved by the ensemble method which evidences
a 10.1% error. This is also the case with GELF for which an error of 15.7% is
manifested. In the case of GELF, the improvements range from 8.0% to 24.9%
compared to all other methods.
Table 4.
Relative Absolute Errors for the Heterogeneous environment
Strategy
RandomR
SVM-RFE
GELF
M5P
22.1%
14.7%
23.7%
ANN
23.7%
11.0%
40.6%
k-NN
19.8%
10.3%
24.3%
SVR
32.6%
25.5%
32.3%
Ensemble model
19.8%
10.1%
15.7%
From the table, similar observations to the homogeneous case are derived.
Higher errors are obtained for the RandomR task and there is a wide mar-
gin between the ensemble model and the remaining ones while modeling the
performance of GELF tasks.
Overall Comparison.
Figure 5 presents the average error for each of the strate-
gies considering the six scenarios. It can be seen that the ensemble strategy
presents the minimum average error (20.6%). These results highlight the robust-
ness of the ensemble model. It is important to note that the standalone M5P
