透過您的圖書館登入
IP:18.224.66.196
  • 學位論文

屬性值隨時間改變的資料分類方法研究

A Classification Using Time-Sequential Attributes.

指導教授 : 陳彥良
若您是本文的作者,可授權文章由華藝線上圖書館中協助推廣。

摘要


隨著資訊科技的進步以及產業電子化與企業整合的趨勢,從龐大的資料中擷取出有價值的資訊,已成為研究與實務上的重要議題,資料挖礦是一種不斷循環的資料分析與決策支援過程,主要是以自動或是半自動的方式從大量資料中探索和分析,以發現出有意義的規則,並將其整理成有價值的知識。   本研究以分類方法的研究為主要重點,過去的分類方法無法針對屬性值會隨時間改變的屬性作分類,因此我們發展出一套新方法,以序列規則的精神為基礎找出所有符合的規則,並建立出分類器,如此就可以對測試的資料做最有效的預測分類。另外,不像一般序列規則決策過程中只建立單一門檻值,由於很多稀有屬性或是類別可能因為門檻值太高而被忽略,或是因為門檻值太低而產生出太多 規則,兩者都是單一門檻值所衍生出的問題,因此我們針對不同屬性以及不同類別分別建立其多重門檻值,以保有稀有但是重要的屬性以及類別;最後,在分類建立後,測試資料的預測方式也因多重門檻值的設立,而發展定義出一套新方法,可以比傳統預測方式更有效且分類正確性更高。   本論文可應用在銀行業或是股市;銀行業的應用中,可以透過顧客多年累積的交易數據、付款情形、繳付信用等有時間順序的資料,以及顧客基本資料等沒有時間順序的資料,共同使用分析,進而給予顧客最適當的類別,可進一步應用在判斷顧客是否合適往後借貸交易,或是其他加銀行決策管理者對顧客做正確判斷的效益,進而幫助降低銀行呆帳成本。同理,也可以利用長時間收集的股市交易資訊等有時間順序關係的資料,透過本研究發展出的演算法,進而針對某股票做分類,分類結果可以提供研究者分析或是預測股市;以上可以看出本研究之應用面很廣泛。

並列摘要


Classification is an important method for class label predicting from databases. Most existing methods, however, assume that attribute-values are all constant. In many real-life applications, however, attribute-values may change at different time, such as the daily stock price, the blood pressure at different time, or others. We call these attributes time-sequential attributes. In this paper, we first extend the traditional classification problem to deal with time-sequential attributes. Next, the algorithm, called MutipleMIS-SP, is presented to generate all classification rules for classifier generation. In our approach, we also consider the concept of multiple minimum supports since each attribute and attribute-value pair doesn’t have similar frequency in the database. Using the concept of single minimum support may lead to rare item problem and finally result in low classification accuracy. Finally, two classification criteria are proposed to predict the class label using the generated classification rules.  Detailed experiments were also presented. Seven synthetic datasets and a real-life dataset, BA-CUSTOMER, were used in our performance analyses and the scalability tests were also given. The result shows that the accuracy of MutipleMIS-SP is better than traditional classification technique C4.5 algorithm in both synthetic datasets and the real dataset.

參考文獻


[40] R. Agrawal and R. Srikant, Fast algorithms for mining association rules., VLDB-94, 1994.
[1] A. Fadlalla, An experimental investigation of the impact of aggregation on the performance of data mining with logistic regression., Information & Management, 42(5), Jul. 2005, pp. 695-707.
[2] B. Lent, A. Swami, and J. Widom, Clustering association rules., ICDE-97, England, April 1997.
[3] B. Liu, W. Hsu, and Y. Ma, Integrating classification and association rule mining., KDD-98, New York, NY, Aug. 1998.
[4] B. Liu, W. Hsu, and Y. Ma, Mining Association Rules with Multiple Minimum Supports, In Proceedings of KDD-99, 1999.

延伸閱讀