A Study of the Ant Colony System for the Discovery of Classification Rules

• Main Author: 陳惠琪
• Other Authors: Duen-Ren Liu
• Format: Others
• Language: zh-TW
• Published: 2006
• Online Access: http://ndltd.ncl.edu.tw/handle/62358540090284351355

碩士 === 國立交通大學 === 管理學院碩士在職專班資訊管理組 === 94 === Ant Colony Optimization (ACO) was proposed by Colorni et al in 1991 from the collaborative behavior of ant colonies. It has been applied to such combinatorial optimization problems as traveling salesman problem, quadratic assignment problem, just to name a few. In the recent years, the ACO approach was deployed in the area of data mining, where algorithmic and statistical techniques are used to discover or extract useful information as well as knowledge from large volume of data. This thesis aims to study the efficiency and effectiveness of Ant-Miner, a well-known classifier that is developed using ACO. The major function of Ant-Miner is to extract classification rules out of the examined data sets. The terms or conditions of a rule will be added or removed by ant colony through collaboration or pheromone sharing. The focus of this research is set on the performance comparison between Ant-Miner and Weka, which is a data mining tool incorporating machine learning mechanisms. In this research, we have two data sets with nominal attributes. The first set is selected from the UCI Machine Learning Repository, and the second is a real data set collected in 2005 by a local research institute. We use the two data sets to compare Naivebayes and Decision Tree with Ant-Miner. Experimental results and analysis show that different classification tools demonstrate different levels of efficiency and effectiveness. We also examine the performance of Ant-Miner resulted from different parameter settings, including such as colony size, evaporation rate and diversification level.