An improved algorithm for oblique decision tree classification based on rough set theory

Oblique decision tree (ODT) is viewed as an important development for decision tree classification. Most of researches are focused on finding an optimal hyperplane for a oblique test, such as regression, linear programming, randomization, simulated annealing and so forth. Empirical studies indicate that these techniques can somewhat improve classification accuracy and reduce tree size compared to the univariate decision tree (UDT). But oblique decision tree have the shortcomings of the complexity in the tree structure expression and over-fitting problem. In this paper, we introduce two concepts, i.e. positive region importance measurement of the condition attributes and generalization thereby combining both to forming a new algorithm-oblique decision tree based on rough set theory, simply RSODT. The algorithm can effectively overcome above drawbacks of ODT. We experimentally test the proposed algorithm in terms of classification accuracy and tree size, using the entire 36 UCI datasets selected by Weka, and compare it with ID3, classification and regression tree (CART), simulated annealing of decision tree (SADT), linear machine decision tree (LMDT), oblique classifier 1 (OC1). The study results show that RSODT algorithm outperforms the comparison classification algorithms with improved classification accuracy and smaller tree size. Copyright © 2011 Binary Information Press.