馬氏-田口系統(Mahalanobis-Taguchi System ; MTS)由田口玄一(Genichi Taguchi)博士於2001年針對多變量資料所提出的分類技術。MTS相異於其他分類方法,其優點主要包含:使用馬氏距離以考慮各特性變數之相關性、資料不需假設符合任何分佈、可用於不平衡(Imbalanced)之數據,以及使用直交表(Orthogonal Array;OA)與SN比(Signal-To- Noise Ratio)進行刪減特性變數以獲得較高的分類準確率。近年來,MTS於業界應用之成功案例包含:富士膠卷(Fuji photo film)、福特汽車(Ford)與富士樂施(Fuji Xerox)等。 由於馬氏距離以橢圓邊界來描述資料集觀測數據以及使用經驗法則或試誤法(Trial and Error)訂定出分類臨界值,而這種量測尺度對於資料描述太過於鬆散。基於上述,本研究提出支援向量資料描述(Support Vector Data Description;SVDD)取代馬氏距離並結合SN比進行特性變數刪減,而發展出SVDD-TS方法。相異於馬氏距離,SVDD以更緊密的邊界來描述資料集觀測數據並運用支援向量(Support Vector)建構出合適資料集之超球體空間,亦即透過支援向量訂定出超球體之臨界值。相較於傳統MTS方法,所提之方法能以客觀方式訂定出分類臨界值(Threshold)。 為了驗證SVDD-TS方法之有效性,本研究將分別使用三個案例:性別資料、行動電話製程案例與院內感染案例來進行驗證。其中性別資料與行動電話製程案例屬於計量型資料(Variable Data)型態;院內感染案例為計數型資料(Attribute Data)型態。結果發現,本研究之SVDD-TS方法確實能有效提升傳統MTS方法之分類準確率。此外,本研究更進一步發現,當資料屬於計數型資料型態時,SVDD-TS方法更提升約13%之分類準確率。
Mahalanobis-Taguchi System is a forecasts diagnostic technology of multivariate data developed by Dr. Genichi Taguchi in 2001. MTS is dissimilar from other classifications . MTS takes advantages of (1)data doesn’t required to meet any distribution;(2) can be used for unbalanced data;(3) use the Signal-To- Noise Ratio to reduce the number of variables and obtain a higher classification detection rate. In recent years, successful of MTS application cases included Fuji Photo Film, Ford, Fuji Xerox. However, Mahalanobis distance build elliptic boundary to describe the samples in dataset, and using the empirical or trial-and-error method to set the classification thresholds. As above mentioned, this study will develop SVDD-TS method via using support vector data description and combine Signal-to-Ratio to reduce the number of variables. SVDD is different to Mahalanobis distance, it will be closer to describe the boundaries of dataset. The use of support vectors can construct suitable data hypersphere space and determine hypersphere threshold. Contrary to MTS, more objective way to set classification thresholds. In order to verify the SVDD-TS method and the MTS method, this research uses three cases: gender data, mobile phone manufacture process, nosocomial infection. Gender data and mobile phone manufacture process belong to variables data type; nosocomial infection is attributes data type. Results showed that SVDD-TS can enhance the traditional of MTS classification accuracy. Furthermore, when data is attributed , SVDD-TS can be enhance nearly about 13% classification accuracy.