|
|
 |
| |
|
|
|
On Post-Processing the Outputs of Prediction Systems: Strategies, Empirical Evaluations and a Case Study in Computer Security |
|
|
|
PP: 1619-1635 |
|
|
|
doi:10.18576/amis/100442
|
|
|
|
Author(s) |
|
|
|
Mouaad Kezih,
Mahmoud Taibi,
Salem Benferhat,
Karim Tabia,
|
|
|
|
Abstract |
|
|
| Supervised classification is a well-known task in data-mining and it is widely used in many real world domains. Classifiers
are automatic prediction systems used to predict the class label of items described by a set of features. In many areas, it is important
to take into account some extra knowledge and constraints in addition to the one learnt or encoded by the classifier. In this paper, we
propose an approach allowing to exploit the available domain knowledge with the predictions of a classifier.More precisely, we propose
to post-process the predictions of a classifier in order to take into account some domain knowledge. This approach can be applied with
any classifier be it probabilistic or not.We propose post-processing criteria and methods to encode and exploit different kinds of domain
knowledge. Finally, the paper provides extensive experimental studies on a representative set of benchmarks and classification problems
including imbalanced datasets.We also provide a case study on two crucial problems in computer security which are intrusion detection
and alert correlation. Interestingly enough, the results show that using only some available knowledge about the training datasets or the
performances of the used classifiers can improve these classifiers’ efficiency while fitting the available domain knowledge. |
|
|
|
|
|
|
|
