Impact of an Extra Layer on the Stacking Algorithm for Classification Problems

Abstract

Classifying and making decisions are tasks performed by any human being in their daily lives. Learning algorithms have been widely studied as tools to aid information management, with an objective to maximize the generalization capacity. Learning algorithms can be used individually or as a committee of machines (ensembles). An ensemble uses the solutions provided by several machines, making different combinations with them to reach a final decision, such as multi-layer algorithm stacking. When combining combination methods, one arrives at a three-layered architecture, which is the focus of this article. The objective of this work was to evaluate the influence of adding one more layer in the stacking meta-learning algorithm in other to obtain accuracy, area under ROC and time in relation to the lower layers, under the influence of the experiment, database and level factors. It was possible to conclude that, statistically, classifiers of the extra layer presented, in a general way, better performance in terms of accuracy and area. However, time grew sharply at each top layer added.