本研究目的主要在於應用超學習(Meta-Learning)增進傳統掛袋法(Bagging)做為預測評估的方法,此方法是根據樣本及預測的結果來組成超學習集,由超學習過程中依據基底分類器的輸出結果進行學習,並使用UCI四個資料庫(德國信用卡資料庫、澳大利亞信用卡資料庫、心臟病資料庫及皮膚病資料庫)做為實驗測試與評估。應用集成式演算法中的掛袋法結合堆疊法(Stacking)與以單一分類器(決策樹、簡單貝氏、支援向量機、倒傳遞類神經網路)和四種方法的組合來進行比較,比較使用和不使用堆疊法是否優於單一分類器。研究結果顯示使用堆疊法優於單一分類器;且使用多種不同基底分類器則會有較佳的準確度。
This study investigates the meta-learning to enhance bagging as prediction method. Meta-learning is based on data sample and prediction results in which the learning process is based on the output of classifiers. The experiment and evaluation of the UCI data sets (Germany credit card information, Australia credit card information, heart disease and skin disease) are tested by meta-learning. We compare the single classifier (decision tree, naive bayes, support vector machine and back-propagation neural network) and the proposed ensemble algorithm with bagging consist of stacking or not. The results indicate that stacking is greater than single classifier and used multiple classifiers can obtained better classification accuracy.