Theory of Generalized Inductive Learning: Induction of Multiple-Decision Outcomes

• Main Authors: Chang Chieh-Li, 張介立
• Other Authors: Wei Chih-Ping
• Format: Others
• Language: zh-TW
• Published: 1998
• Online Access: http://ndltd.ncl.edu.tw/handle/34456366917483532278

碩士 === 國立中山大學 === 資訊管理研究所 === 86 === Current business and organizations are facing rapid changes in environment. They require more knowledge or higher level information tosupport decision making and gain strategic advantages. The requiredknowledge or higher level information is often hidden in business and organizations' data asset. The process of extracting useful informationwhich is previous unknown, valid and actionable for making crucialbusiness decision from data asset is called "data mining". According tothe function they perform, data mining techniques be grouped into predicting modeling, cluster analysis, link analysis and deviation detection. Of the predicting modeling, there are two specializations,classification and value prediction. Among all data mining techniques,classification is the most common data mining task in business now. Although classification is widely employed by organizations, existingclassification techniques are still limited to handle complexapplications. When applications involve multiple decision outcomes,existing classification techniques can not directly be applied. In thisresearch, we first overviewed such techniques as ID3, CN2, Decision Class Revision (DCR), Multiple-Decision-Tree Induction (MDTI), and backpropagation. Then, we proposed a new induction system, calledGeneralized Decision Tree Induction (GDTI), which is capable of handling multi-decision-outcome problems. Besides the multi-decision outcomes, GDTI is also suitable to single-decision-outcomeapplications. An evaluation on GDTI shows that the time complexity is not as satisfactory as such techniques as DCR and MDTI. But we believe thatthis is a tradeoff between learning performance and time execution. Anempirical evaluation showed that GDTI is comparable to a well-trainedbackpropagation network.