Interaction of a Superscalar Architecture and a Data Mining Application

As problem sizes continue to grow there is more and more interest in keeping the run times reasonable. Superscalar processors offer the opportunity to increase performance via parallelism in the hardware. Compiler technology has tried to help the user take advantage of this parallelism by restructuring the code to make it run more efficiently on modem processors. Although there has been tremendous success in compilers, there are limits as to what can be done without a deeper understanding of the algorithms used, implementation and runtime behavior. This paper looks at the aspects of decision tree construction. It focuses on C4.5 , which is a popular decision tree generator. First, classifiers in general are discussed. The difference between deductive and inductive reasoning is then considered. A model is introduced that is used to formulate the problem of classification. Shortcomings of the classical statistical method of n-way analysis of variance are examined and an overview of the general concepts of decision trees with details of the implementation of C4.5 is presented. Finally, a section of C4.5 is examined to illustrate an area where modifications to the code could result in greater utilization of the parallelism available in the hardware.