Chapter Summary

The results of our experiments with BruteDL reveal:

• BruteDL provides satisfactory performance when relatively well trained on many data sets.
• The default rule gives rather accurate prediction when the data are not evenly distributed among the possible classes, but it does not perform as well when the training data are evenly distributed among the classes, the default rule leads to most noticeable degradation of performance when its coverage is higher and the classes are more equiprobable.
• The prediction accuracies of the default rule on the data sets with roughly even class distributions are below 50%, which could be viewed as poor. The performance of the whole system on most of the data sets is still satisfactory only because the default rule does not classify many of the testing observations, and the other rules learned by BruteDL perform very well.

The next chapter explores an alternative to the default rule as default strategy.