Affiliations: [a] School of Information Science and Technology, Southwest Jiaotong University, Chengdu, Sichuan, China | [b] Department of Information Engineering and Computer Science, Feng Chia University, Taichung City 40724, Taiwan | [c] Department of Computer Science and Information Engineering, Asia University, Taichung 41354, Taiwan
Correspondence:
[*]
Corresponding author: Xu Zhuang, School of Information Science and Technology, Southwest Jiaotong University, Chengdu, Sichuan, China. E-mail:zhuangxu6620362@163.com
Abstract: In empirical data modelling, a model of system is built up from a set of
cases that the system has observed. Eventually, the performance of the
inducted model is dominated by the quality and quantity of observations.
Feature transformation methods are widely used to improve quality of
knowledge extracted from observations to build up more accurate and robust
model. In the paper, a new feature transformation method named dynamical
feature bundling for decision tree algorithm is proposed. Dynamical feature
bundling groups a set of features in the tree induction phase and it enables
decision tree algorithms to 1) make use of features in one bundle together
to make collective judgments in splitting phase; 2) learn more reliable and
stable knowledge from feature bundles created based on domain knowledge of
experts; 3) embed feature transformation step into tree induction phase, and
therefore the extra pre-process step which are necessary for static feature
transformation methods is inessential. Our experiments show 2%-9%
improvements of AUC value on a very imbalanced dataset. Slight improvements
are also obtained on a more balanced data set.
Keywords: Decision tree, feature bundling, feature transformation, web spam detection
