標題: 利用距離轉換於以配對距離為基礎之機器學習
Applying Distance Transformation to Pairwise Distance-based Machine Learning.
作者: 游旻哲
Yu, Min-Che
胡毓志
Hu, Yuh-Jyh
生醫工程研究所
關鍵字: 分類法;以距離為基礎的學習演算法;距離轉換;classification;(dis)similarity-based learning;distance transformation
公開日期: 2013
摘要: 於真實世界中,有許多資料是以距離亦或是相似度來表示資料點彼此之間的關係。 常見的定理和學習演算法大多數皆以向量為基礎所設計,並且有很好的解釋和應用於向 量型態的資料。然而,以距離為基礎的定理和學習演算法則相較於稀少,而且於資料型 態的轉換上,以向量型態所表式的資料轉換成以距離型態來表式資料明顯容易於以距離 型態的資料轉換成向量型態的資料。因此,本論文提出一種以距離為基礎的轉換方法, 使用隨機的機制與迭代的模式來修改資料點彼此之間的距離,達到不相同類別之間的邊 界清楚化。我們與常見的八種學習演算法來進行實驗。不管是在距離型態的資料集還是 向量型態的資料集之實驗,我們提出的距離轉換方法所得的準確率皆比其他八種常見的 學習演算法來的高。最後會以實驗來探討與證明距離轉換方法是有效的且可行的。
Relationships among instances are typically represented by (dis)similarities in the real world. Many theories and algorithms have been developed to solve learning problems of vectorial data by seeking the hypothesis that best fits the observed training sample. However, (dis)similarity-based learning algorithms are scarcer than vectorial-base learning. It is relatively easier to convert vectorial data into (dis)similarity data than vice versa. Therefore, we propose a distance-based transformation method for similarity-base learning. It reveals a clearer class boundary implied in data by modifying the (dis)similarities between examples. We compare performace with eight classifiers. To demonstrate its performance, we compared it with eight other learning algorithms in both (dis)similarity data and vectorial data. The experimental results show that the proposed method is better than (or comparable to) the other classifiers in prediction accuracy.
URI: http://140.113.39.130/cdrfb3/record/nctu/#GT070056707
http://hdl.handle.net/11536/74003
顯示於類別:畢業論文