Associating kNN and SVM for pattern recognition

淡江大學機構典藏 > 工學院 > 機械與機電工程學系暨研究所 > 學位論文 > Item 987654321/35519

請使用永久網址來引用或連結此文件: https://tkuir.lib.tku.edu.tw/dspace/handle/987654321/35519

題名:	Associating kNN and SVM for pattern recognition
其他題名:	聯結k最近鄰域與支向機之模式識別
作者:	許哲彰;Hsu, Che-chang
貢獻者:	淡江大學機械與機電工程學系碩士班楊智旭;Yang, Jr-syu
關鍵詞:	支持向量機;k最近鄰域法;損失函數;模式識別;Support Vector Machine(SVM);k-Nearest-Neighbor(kNN);Loss Function;Pattern Recognition
日期:	2006
上傳時間:	2010-01-11 06:43:56 (UTC+8)
摘要:	本論文建立一個改良型的模式識別方法，它結合無母數k最近鄰域法(kNN)於支持向量機(SVM)用來辨識情勢特殊而又深具義意的訓練樣本，尤其是座落於兩類別重疊區域間難以辨識的樣本。本方法區分為兩階段來進行模式識別，首先使用k近鄰域加強器來過濾篩選上述類型樣本，然後根據這些過濾後的資料來進行支持向量機的分類，因此在兩階段的分類的過程中，將形成較大的懲罰值於這些需要技術性才能加以區分的困難樣本，來獲取較嚴峻的損失函數，而這樣的處裡確實在系統的最佳化過程產生不錯的效果。在第一階段使用k最近鄰域加強器來收集訓練樣本的資訊並且產生一系列的權重值給予每一筆樣本，然後在第二階段時，由支持向量機使用參數化實數值類別標籤來攜帶各種不同層次權重的訓練樣本資訊以做分類。因此依照這種改良型分類器的演繹程序所得到的決策邊界被用來引導一個較為精準的分類狀況。採用k近鄰域加強器的緣故可在估算區域密度時所展現的優勢得知，也就是對於那些被訓練集所遮蔽的樣本點，能夠僅只考量它們本身所散落的區域位置即可。這樣的模型能夠提供一種用來強調被忽視的小區域孤立樣本以及難以區隔的樣本，也就是說主要就是用來處理這些需要較多技術性指導才能加以區隔的樣本。經由數值模擬結果可知，本方法確實可以改善被訓練集所遮蔽的樣本點以及當系統受到參數擾動下仍保持良好的運算特質。 This thesis conceived a new model merging a non-parametric k-nearest-neighbor (kNN) method into an underlying support vector machine (SVM) to emphasize the meaningful training examples, especially for the terms of the difficult examples which were located on the overlapping region of the training set. The model consisted of a filtering stage of kNN emphasizer and a classical classification stage. The model motivated by adding heavier penalties into the difficult examples to attain a stricter loss function for optimization could really take effect on emphasizing the difficult examples. First, the filtering stage of the kNN emphasizer was employed to collect information from training examples and to produce a set of emphasized weights for each example. Then, an underlying SVM with parameterized real-valued class labels was used to carry the information in various confidence levels to the training examples for classification. The novel idea of real-valued class labels for conveying the emphasized weights permitted an effective way to pursue the optimal solution inspirited by the additional information. A slight alteration of resultant decision boundary was therefore produced after the convex programming of the SVM, and conducted a higher accurate classification in the training. The adoption of the kNN emphasizer did make sense due to the local density estimation which had the advantage to screen out the difficult examples among the training set by considering merely the distinct situation of the examples themselves. The proposed model provided a way for emphasizing the substantial and subtle examples in the learning process, especially for the difficult examples. Moreover, discussions with detailed experimental evidence were also presented in the paper category to address the corresponding properties. All of those properties were consistent in both the theoretical derivations and related experimental results.
顯示於類別:	[機械與機電工程學系暨研究所] 學位論文

文件中的檔案:

檔案	大小	格式	瀏覽次數
	0Kb	Unknown	530	檢視/開啟

在機構典藏中所有的資料項目都受到原著作權保護.

TAIR相關文章

資料載入中.....