done

report/main.md

@@ -419,5 +419,14 @@ test is inconclusive.
 
 ## Practical Usefulness
 
-Discuss the practical usefulness of the obtained classifiers in a
+The evaluation shows that all classifiers but the Naive Bayes classifier outperform the biased classifier in overall performance
-realistic bug prediction scenario (1 paragraph).
+(represented by the F1 score). Given the evaluation was not performed on a dedicated test set, all classifiers may be subjected to
+training bias, which might yield better metrics than a similar evaluation performed on completely new data. However, given the 
+evaluation sample size (i.e. the number of training runs) this effect is minimal.
+
+Given this premise, I can say with reasonable confidence the *DT*, *MLP*, *RP* and *SVP* 
+classifiers would be useful in predicting potential bugs in the Google JSComp project, i.e. the source of the used dataset. Given the 
+literature presented during lecture, it is not certain if the same trained classifier would yield acceptable results for source
+code from another project. This is because differences in coding conventions, the bug tracking process, or simply the definition of 
+what constitutes a bug may vary from project to project. A solution to this problem might be to simply train the classifiers on a dataset
+coming from the project where bug prediction is needed.