Decision tree based classification algorithms like C4.5 and Explore build a single tree from a data set. The two main purposes of building a decision tree are to extract various patterns/logic-rules existing in a data set, and to predict the class attribute value of an unlabelled record. Sometimes a set of decision trees, rather than just a single tree, is also generated from a data set. A set of multiple trees, when used wisely, typically have better prediction accuracy on unlabeled records. Existing multiple tree techniques are catered for high dimensional data sets and therefore unable to build many trees from low dimensional data sets. In this paper we present a novel technique called Sys- For that can build many trees even from a low dimensional data set. Another strength of the technique is that instead of building multiple trees using any attribute (good or bad) it uses only those attributes that have high classification capabilities. We also present two novel voting techniques in order to predict the class value of an unlabeled record through the collective use of multiple trees. Experimental results demonstrate that SysFor is suitable for multiple pattern extraction and knowledge discovery from both low dimensional and high dimensional data sets by building a number of good quality decision trees. Moreover, it also has prediction accuracy higher than the accuracy of several existing techniques that have previously been shown as having high performance.
|Cite as: Islam, Md. Z. and Giggins, H. (2011). Knowledge Discovery through SysFor - a Systematically Developed Forest of Multiple Decision Trees. In Proc. Australasian Data Mining Conference (AusDM 11) Ballarat, Australia. CRPIT, 121. Vamplew, P., Stranieri, A., Ong, K.-L., Christen, P. and Kennedy, P. J. Eds., ACS. 195-204 |
(local if available)