分類是依據已知的資料及其類別屬性來建立資料的分類模型，並以此預測其他未經分類資料的類別，是一項應用非常廣泛的資料探勘技術。其中決策樹是最常使用的一種分類技術，因為它有容易了解、計算效率高的特性。決策樹廣泛使用在訊號分析、專家系統、醫療辨識等領域裡。但決策樹常因為訓練資料內含的雜訊資料、特殊案例的影響，造成樹體結構龐大、分支太多，產生規則過多難以理解與應用的問題，此項缺點減少了決策樹的可用性。 因此本研究透過限制決策樹的葉節點數，控制決策樹產生的規則量，並在使用者給定的葉節點數範圍內，達到最高的準確度。我們發展出一套新的演算法，本演算法以階層式分群法中的聚合法合併決策樹的分支，限制決策樹為二元樹，以便控制決策樹的節點數量。最後本研究再以實際資料進行實驗實作。 實驗結果顯示，我們提出的新演算法與C4.5比較，在同樣的葉節點數限制下，達到比C4.5更好的準確度。 Classification, which builds a data classification model based on attribute value and label of existing data, is a very widespread data mining technology. Decision tree is one of the most popular classification technologies, because it is easy to understand and has the high efficiency computing. Decision tree is widely applied to signal classification, expert system, and medical diagnosis. Because of the noise data and special case of training data sets, decision tree is always huge and it contains too many branches and rules which are difficult to understand. This shortcoming reduces the availability of decision tree. Therefore, we reduce rules from a decision tree by limiting the number of leaf nodes of the decision tree and achieve the highest accuracy with the number of leaf nodes given by user. For this purpose, we propose a new algorithm. We use the agglomerative approach of the hierarchical clustering to limit the decision tree to binary tree by combining the branches of decision tree. Experiment results show that compared with the C4.5, the proposed algorithm successfully reduces the number of leaf nodes and makes better accuracy.
