Data Classification Based On Ant Colony Classifier

Many problems in reality can be transformed into classification problem indata mining. This makes the classification technique may be applied to many areas,such as commercial insurance, network measurement, weather forecast,bio-informatics. However, the data accumulated in various areas is graduallyinflated with the continuous development of IT. Therefore, how to improve theoperating efficiency of the classification technique is the key to solve massive dataclassification problem effectively.The Ant-Miner is comparable to classic classification algorithms, which isbased on Ant Colony Optimization in Swarm Intelligence. However, Ant-Minerdoes not make full use of the idea of the ant colony, and its heuristic strategycontains local information. Ant-Miner is different from ACO, because the value ofAnt-Miner heuristic function is modified during the running of the algorithm, andthis increases the computational complexity of the algorithm. For above problems,the main contributions of this thesis are as follows:(1) The definition, basic principles and the implemented algorithms of dataclassification are reviewed. The ideological source, basic way of working and coresteps of Ant Colony algorithm are introduced. The basic principles of Ant-Minerand the research progress of methods for improving it are described.(2) In order to improve the efficiency of Ant-Miner solving the problem ofdata classification, an improved ant colony classification algorithm namedmAnt-Miner+is proposed. This algorithm draws on the idea of Ant-Miner usingmulti-ants, and introduces a new heuristic strategy. Experimental results on UCIdatasets show that, mAnt-Miner+improves the operational efficiency withoutinfluence on the prediction accuracy and the simplicity of rule lists, and overcomesthe premature convergence problem of mAnt-Miner.(3) The proposed mAnt-Miner+is applied to the membership classification ina retail company. Experimental results on member datasets show that, in terms ofrunning efficiency, mAnt-Miner+is more efficient than mAnt-Miner and Ant-Miner;in terms of prediction accuracy and the simplicity of rule lists, mAnt-Miner+iscompetitive with Ant-Miner and mAnt-Miner; however, mAnt-Miner+overcomesthe instability of mAnt-Miner. Through the analysis of discovery rules, somevaluable information to retail company is obtained.