| dc.description.abstract | While there are many approaches to data mining, it seems that there is a hole in the ability to make use of the advantages of multiple techniques. There are many methods that use rigid heuristics and guidelines in constructing rules for data, and are thus limited in their ability to describe patterns. Genetic algorithms provide a more flexible approach, and yet the genetic algorithms that have been employed don't capitalize on the fact that data models have two levels: individual rules and the overall data model. This dissertation introduces a multi-tiered genetic algorithm capable of evolving individual rules and the data model at the same time. The multi-tiered genetic algorithm also provides a means for taking advantage of the strengths of the more rigid methods by using their output as input to the genetic algorithm. Most genetic algorithms use a single "roulette wheel" approach. As such, they are only able to select either good data models or good rules, but are incapable of selecting for both simultaneously. With the additional roulette wheel of the multi-tiered genetic algorithm, the fitness of both rules and data models can be evaluated, enabling the algorithm to select good rules from good data models. This also more closely emulates how genes are passed from parents to children in actual biology. Consequently, this technique strengthens the "genetics" of genetic algorithms. For ease of discussion, the multi-tiered genetic algorithm has been named "Arcanum." This technique was tested on thirteen data sets obtained from The University of California Irvine Knowledge Discovery in Databases Archive. Results for these same data sets were gathered for GAssist, another genetic algorithm designed for data mining, and J4.8, the WEKA implementation of C4.5. While both of the other techniques outperformed Arcanum overall, it was able to provide comparable or better results for 5 of the 13 data sets, indicating that the algorithm can be used for data mining, although it needs improvement. The second stage of testing was on the ability to take results from a previous algorithm and perform refinement on the data model. Initially, Arcanum was used to refine its own data models. Of the six data models used for hypothesis refinement, Arcanum was able to improve upon 3 of them. Next, results from the LEM2 algorithm were used as input to Arcanum. Of the three data models used from LEM2, Arcanum was able to improve upon all three data models by sacrificing accuracy in order to improve coverage, resulting in a better data model overall. The last phase of hypothesis refinement was performed upon C4.5. It required several attempts, each using different parameters, but Arcanum was finally able to make a slight improvement to the C4.5 data model. From the experimental results, Arcanum was shown to yield results comparable to GAssist and C4.5 on some of the data sets. It was also able to take data models from three different techniques and improve upon them. While there is certainly room for improvement of the multi-tiered genetic algorithm described in this dissertation, the experimental evidence supports the claims that it can perform both data mining and hypothesis refinement of data models from other data mining techniques. | |
