A Study of Random Linear Oracle Ensembles
Random Linear Oracle (RLO) ensembles of Naive Bayes classifiers show excellent performance . In this paper, we investigate the reasons for the success of RLO ensembles. Our study suggests that the decomposition of most of the classes of the dataset into two subclasses for each class is the reason for the success of the RLO method. Our study leads to the development of a new output manipulation based ensemble method; Random Subclasses (RS). In the proposed method, we create new subclasses from each subset of data points that belongs to the same class using RLO framework and consider each subclass as a class of its own. The comparative study suggests that RS is similar to RLO method, whereas RS is statistically better than or similar to Bagging and AdaBoost.M1 for most of the datasets. The similar performance of RLO and RS suggest that the creation of local structures (subclasses) is the main reason for the success of RLO. The another conclusion of this study is that RLO is more useful for classifiers (linear classifiers etc.) that have limited flexibility in their class boundaries. These classifiers can not learn complex class boundaries. Creating subclasses makes new, easier to learn, class boundaries.
KeywordsClassifier Ensemble Naive Bayes Clusters Subclasses
Unable to display preview. Download preview PDF.
- 11.Rish, I., Heellertein, J., Jayram, T.: An Analysis of Naive Bayes on Low-Entropy Distributions, Tech. Report RC91994, IBM T. J. Watson Research Center (2001)Google Scholar
- 12.Rodriguez, J.J., Kuncheva, L.I.: Naive bayes ensembles with a random oracle. In: Haindl, M., Kittler, J., Roli, F. (eds.) MCS 2007. LNCS, vol. 4472, pp. 450–458. Springer, Heidelberg (2007)Google Scholar
- 14.Vilalta, R., Achari, M.R., Eick, C.F.: Class Decomposition via Clustering: A New Framework for Low Variance Classifiers. In: ICDM 2003 (2003)Google Scholar