Bucket Learning: Improving Model Quality through Enhancing Local Patterns
Guangzhi Qu, Hui Wu · 2009
It is always desirable to improve the quality of a global classification model with the existence of other models. In this work, bucket learning methodology is first proposed to improve the model quality through enhancing its local patterns. We formally define the concept of slab as a tri-tuple, which unifies the data view, model view and evaluation view of a data mining task. The bucket learning framework includes main modules of slab generation, short slab discovery, and short slab replacement as necessary steps to improve the model's quality. Algorithms are designed to facilate the operations of quantifying the model merits, identifying the inferior local patterns and improving the global model. A prototype system is developed to verify the proposed methodology. The bucket learning prototype system is evaluated on 16 representative data sets from UCI data repository. Experimental results show that the improved model have an averaged F-measure of 79.5% with an improvement of 7.3% from the original model learned by J48.