多階段調適樹群模型建構方法及其於腫瘤分級之應用

Abstract

在二類別分類問題上，諸多研究已驗證分類樹群(Ensemble)模型的分類效能比單一分類樹來得良好；然而，在實際應用的領域中，僅有兩種類別的分類結果往往不足以滿足實際需求，例如在臨床腫瘤診斷上，醫師對腫瘤所做的判斷不僅僅只有良性或惡性如此非黑即白的結果，而是針對腫瘤的各項特徵，評估腫瘤為良性或惡性的機率高低。 再者，分類樹群雖有較好的分類效能，卻不如單一分類樹容易解釋。在單一分類樹中，我們可由一案例所經過的節點與分枝來了解該樣本被分類為某一類別的原因，但在分類樹群模型裡，一案例會經過多棵分類樹，並由此眾多分類樹共同判定該案例的類別，如此一來，該樣本之所以被歸為某一類別的原由便隨之複雜許多。 針對二類別分類不足以滿足實際需求的問題，雖已有學者提出可將樹群模型分類結果轉換為機率值，提供了比二類別分類更細緻的多類別分類能力，但要適切地處理多類分類問題並不是如此一蹴可及，以一般建構方式建立的樹群模型仍有其不足之處；因此本研究提出階段性地建構樹群模型，並透過參數的設定，使不同階段的樹群模型各自具有其所專精的分類目標，以達到更良好的多類別分類效能。而針對不易解釋的缺點，本論文則提出以樹葉節點的摘要方式對每個案例之所以被歸類為某一類別的原因提供說明。

Literature has shown that ensemble model usually out-performs one single classification tree for binary classification; however, classifying only two classes often fails to satisfy real practice. Take nodule diagnosis for example, what a clinician does is judge how possible a nodule is benign or malignant by the characters of the nodule instead of simply give a deterministically binary answer such as benignancy or malignancy. Moreover, though Ensemble model has better performance, it’s not straightforward to interpret how the prediction is made. Such a question can be easily answered when one single classification is used; the reason lies in the nodes and branches an instance passes by in the classification tree. While in an ensemble model, an instance is judged by a bunch of classification trees, which means there will be far more nodes and branches an instance passes by, and those nodes and branches belong to different trees. Some researchers had provided methods that can transform ensemble prediction into probability and make ensemble be able to deal with more sophisticated problem other than binary classification. However, an ensemble model built by the traditional way is still not capable enough to well perform multi-class tasks such as nodule categorization. Therefore, in this research, a stage-wise ensemble modeling method is offered to build a compound ensemble model where each ensemble component, with a parameter, has its own strength to classify certain subjects. Besides, for ensemble model’s shortage of being difficult to be explained, this research also develops a leaf nodes summarizing method that can provide reasons why an instance is judged to be a certain class.