Meta-learning enhancements by data partitioning

In data mining, there is no learning algorithm which attains the highest accuracy on any dataset. Multilevel arbiter and combiner arbiter are presented in this paper, as techniques to integrate classifiers induced from partitioned data, having as optimization criterion the accuracy of a given dataset. Experimental evaluations have shown that an arbiter tree can be found having similar or higher predictive performance when compared to the accuracy of the individual learner, trained on the entire training set. Moreover, for multiclass dataset with unbalanced class distribution, the combiner arbiter strategy yielded a good improvement in the prediction performance level.