傳統的數值資料挖掘上，證券資料大都是與固定的絕對數值做一比較，(例如：KD隨機指標中的D > 85或K < 20)，而股市資訊間相對的比較鮮少被討論到(例如：隨機指標中K與D 之間的關係)，因此本研究提出一個數值型資料的延伸比較架構，此架構除了基本的「絕對關係比較」外，另外增加了數值間的「相對關係比較」，藉此得知數值間的大於小於等於關係，之後再進一步利用簡單易懂的C5.0決策樹分類方法，拓展連續型數值的比較關係，在絕對與相對關係為基礎下，能挖掘出數值間分類的界限是什麼，找出資料間進一步的「變動關係比較」。 本研究提出有別於傳統絕對比較之相對資料挖掘方法，改善了決策樹在單純絕對數值關係上的限制，針對資料中的關係比較做進一步的探討與實驗，主要將資料分為三種關係，包括了：絕對關係比較、相對關係比較以及變動關係比較。 在基本的「絕對關係比較」上，本研究引進了「相對關係」與「變動關係」的比較，將證券交易日的歷史資料，透過各組實驗數據與統計t檢定來驗證，證明了「絕對+相對關係」較「絕對關係」有較高的學習正確率與精確率，而「變動關係」相較於「絕對+相對關係」也有較高的學習正確率。因此，本研究架構除了能表達傳統決策樹資料挖掘的基本絕對比較概念外，能挖掘出其他更多元、更豐富的相對與變動比較規則，找出具有潛在價值的概念與更完整的數值關係。 In the traditional numerical data mining, the stock data is usually compared with fixed value (ex: the stochastic indicator D > 85 or K < 20). The relative comparison between stock information was rarely discussed (ex: the relation between K and D). Thus we propose an extended comparative framework on the numerical data. This framework includes the basic comparison “absolute comparison”. Besides, the “relative comparison” between values is added. The “greater than” and “smaller than” relationship will be obtained then. To advance further, this thesis makes use of understandable C5.0 decision tree classification method. In addition to “absolute comparison” and “relative comparison”, the “variable comparison” of values boundary would be found. We propose a different framework on data mining method which improves the decision tree to deal with each comparison and do some researches on data comparisons. In this thesis, there are three data types of comparison, and these are: absolute comparison, relative comparison, and variable comparison. We propose “relative comparison” and “variable comparison” for basic “absolute comparison”. As the result of t test via experiments, the accuracy and precision rate of “absolute + relative comparison” is higher than “absolute comparison”, and the performance of “variable comparison” is better than “absolute + relative comparison” significantly. Hence, this framework not only represents the basic “absolute comparison” of traditional data mining but also discovers diversified “relative comparison” and “variable comparison”. In this framework, potential valuable concept can be found.
