Bucket Learning: Improving model quality through enhancing local patterns

Abstract

It is always desirable to improve the quality of a global classification model in the light of the existing models. In this work, the Bucket Learning methodology is first proposed to improve the model quality by enhancing its local patterns. We formally define the concept of a board as a tri-tuple 〈 D, M, R〉, which unifies the data view, model view and evaluation view of a data mining task. The Bucket Learning framework includes the modules of Boards Generation, Short Boards Discovery, and Short Boards Replacement. A prototypical system is developed to verify the proposed methodology. The experimental results on eight representative data sets from the UCI data repository show that Bucket Learning performs better than traditional classification methods such as J48, AdaBoost, Bagging and LogitBoost. We also demonstrate that the Bucket Learning framework can combine all kinds of data classification models and that the combined model outperforms each individual one.