Comparing FBTSeg and NNTree Implementations with Established Ensemble Methods

This paper compares implementations of FBTSeg, a recent experimental segmentation method, and of NNTree, originally a neural network tree based segmentation method, with traditional methods for the combination of classifiers, namely bagging, boosting, and the traditional segmentation using information gain to split. The tests were carried out using three data mining techniques with distinct characteristics, specifically, linear regression, logistic regression, and multilayer perceptron neural networks, in four artificially built datasets. The datasets design was aimed at understanding the specific circumstances where each method, simple or combined, would present better performance. Results from this experiment suggest that blending classifiers using segmentation is a viable solution to improve the performance of both statistic regressions, that both FBTSeg and NNTree are in general more predictive than the traditional segmentation, while bagging and boosting are more effective alternatives for improving neural networks models.