藉由過去的知識與經驗來預測未來事件發生趨勢是管理的先決條件,預測未來事件主要使用分類機;分類機模式是將數據經過適當前處理程序後,並將其資料訓練分類機,然後進行事件預測。前處理程序包含資料填補、極端值偵測、資料離散化、特徵選取四個步驟;由於,不同的前處理程序與不同的分類機的搭配組合(亦可稱之為分類決策程序)會影響分類的準確度;過去文獻,經常使用嘗試錯誤的方法決定分類決策程序,然而,如此確會耗費大量的人力作業成本且並無法保證獲得最佳分類決策程序;本研究使用啟發式演算法決定近似最佳化分類決策程序,並利用UCI資料庫的資料驗證本方法的有效性;最後說明分析結果、方法限制與未來可能研究加以總結。
Using knowledge and experiments to predict the trend of future events is the prerequisites of management. Classifiers are the main models used to predict future events. Data processed by preprocessors are employed to train classifiers and generate information for predicting future events. In this study, the data preprocessor includes four steps (1) data imputation (2) outlier detection (3) data distretization (4) feature selection. Different classifiers are suitable for different data preprocessors; and this procedure can be treated as a classification decision process. The classification decision process influences the classification accuracy. In previous literature, experts usually used trial and error method to determine the classification decision process. However, the trial and error process is time-consuming and can not guarantee to obtain the best classification decision process. This study uses meta-huristics to yield the near optimal classification decision process. Some data in UCI library were used to demonstrate the performance of proposed method. Finally, the experimental results, limitations of proposed method and future research directions were presented.