Information Technology Reference
In-Depth Information
Fig. 4.
Temporal behavior of SRABNET when applied to a dynamic environment,
with more and more samples arriving along time
done as shown in Fig. 4. This dataset was used by Knidel et al. in [14] to illustrate
the robustness of the algorithm on classes with non-convex distributions, but in
this case we are trying to illustrate the temporal behavior of the messages. In
this pictorial example, the messages are being described solely by two numerical
attributes. To reproduce the behavior of SRABNET in Fig. 4, in what we call a
dynamic environment, we determined a sequence of steps. In each iteration the
algorithm takes samples to from all the previous steps to training and test on the
next step samples of the dataset. Each time a new test is performed, the training
set grows with the addition of the previous step samples and the algorithm is
retrained. Intuitively , we can realize that the larger the dataset, the lower the
value of the error rate on the test data. However, the generalization capability of
the model to unseen samples reduces if it becomes too specialized (overtrained).
In the context where the data changes over time, a good model is the one
that can track the changes of distribution, or in this case the change of concept,
