Information Technology Reference
In-Depth Information
Comparison of recall
100
80
AdaBoost M1
BrownBoost
AdaBoostHyb
60
40
20
0
Databases
Fig. 3.2.
Rate of recall
important error rates. Based on this results, we use test-student and we find a
significant p-value 0.0010. We have a significant average gain of 4.8 compared to
AdaBoost.
Considering Brownboost, we remark that it improves the recall of Ad-
aBoostM1, overall the data sets (except the TITANIC one). However, the recall
rates given by our proposed algorithm are better than those of BrownBoost.
Except, with the zoo dataset. In this case, we have a significant p-value 0.0002
and not a significant average gain (1.4)compared to AdaBoost according to the
results given by AdaBoosthyb.
It is also noted that our approach improves the recall in the case of the
Lymph base where the error was more important. It is noted though that the
new approach does not act negatively on the recall but it improves it even when
it can not improve the error rates.
3.4.3
Comparison with Noisy Data
In this part, we are based on the study already made by Dietterich [6] by adding
random noise to the data. This addition of noise of 20% is carried out, for each
one of these databases, by changing randomly the value of the predicted class
by another possible value of this class.
Graphic 3 shows us the behavior of the algorithms with noise. We notice
that the hybrid approach is also sensitive to the noise since the error rate in
generalization is increased for all the databases.
