Feature selection and learning

We did not do any feature selection, other than that built into the two algorithms. We considered two different learners, but in adherence with the rest of the system, we found that a per-class technique was practically most effective. Thus for each class, we created a binary learner. In addition, each learner could only use synthetic events generated by the clustering of the instances of that class. For example, a cluster from the class bad was not used as an attribute for learning the definition of the class come. Thus for each class, the learner was used to build a binary classifier that returned true if the instance was a member of the class and false otherwise.

The two learners we considered were:

• A naive Bayes learner [Mit97] - a learning technique which assumes that all the attributes are conditionally independent and makes an estimate of the probability that a particular training instance belongs to the class as the product of the probabilities that the each value came from the class. Each class learner predicted the probability that the test instance was a member of the class. The binary classifiers were combined by looking for the classifier that returned the highest probability that the instance was a member of its class.
• C4.5 [Qui93], a decision tree builder. Each class learner was asked to classify whether each test instance belonged to it or not. The classifiers produced by this technique were combined in the following way: If no binary classifier ``claimed'' the test instance, it was unclassified. If one binary classifier claimed the test instance, it was classified as that class. If more than one binary classifier ``claimed'' the test instance, the classifier that claimed to have the lowest error rate got it. In addition, the decision trees generated by C4.5 were analysed and considered to see if they produced intelligible results.