Does more randomness help increasing diversity in ensemble learning?

作者：Xiao Zhibo; Li Ping; Fu Yingliang; Lu Mingyu^*

来源：Journal of Convergence Information Technology, 2012, 7(19): 309-318.

DOI：10.4156/jcit.vol7.issue19.37

• Accuracy measures
• Base classifiers
• Data sets
• Diversity
• Diversity measure
• Ensemble learning
• Experimental parameters
• Large scale experiments
• Random projections
• Randomness

The relationship between diversity and accuracy has always been a central issue in ensemble learning research. In this paper, we conducted extensive experiment, hoping to investigate the relationship among randomness, diversity and accuracy in ensemble learning. We first perform largescale experiments to make clear the experimental parameters, including base classifier type, accuracy measure, and diversity measure. Then using Random Decision Trees as the basic base classifier, we tested the effective of more randomness in ensemble learning. Furthermore, we use random projection technique to inject even more randomness into the system, as the experiments on UCI datasets and real-world dataset show that more randomness/diversity indeed can get higher accuracy, but so much more randomness will harm the accuracy performance.

• 出版日期2012