複數模糊類神經系統於多類別分類問題之研究; A Study on Multi-class Classification Using Complex Neuro-Fuzzy System

NCUIR > School of Management at National Central University > Graduate Institute of Information Management > Electronic Thesis & Dissertation > Item 987654321/61097

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/61097

Title:	複數模糊類神經系統於多類別分類問題之研究;A Study on Multi-class Classification Using Complex Neuro-Fuzzy System
Authors:	羅偉成;Lo,Wei-Cheng
Contributors:	資訊管理學系
Keywords:	多類別分類問題;複數模糊類神經系統;一對全部;混合式學習法;F-score;屬性選取;標準粒子群演算法;遞迴最小平方估計法;multi-class classification;complex neuro-fuzzy system (CNFS);one-against-all;hybrid learning;F-score;feature selection;standard particle swarm optimization;recursive least squares estimator
Date:	2013-07-12
Issue Date:	2013-08-22 12:11:48 (UTC+8)
Publisher:	國立中央大學
Abstract:	本研究提出一種分類器架構：CNFS-OAA，其為一基於複數模糊類神經系統 (Complex neuro-fuzzy system, CNFS) 的建模程序，透過一對全部(One-against-all, OAA) 方法將資料集分解為多個二類別資料，並以動態探勘模糊法則的方式來處理分類問題。在CNFS的建模過程中，將使用標準粒子群演算法(Standard particle swarm optimization, SPSO) 來調整其前鑑部參數，與遞迴最小平方估計法(Recursive least squares estimator, RLSE) 來調整其後鑑部參數。而CNFS的法則探勘方式，其法則數量將依據訓練階段之分類正確率來動態增加。當訓練正確率未達到門檻值時，將會探勘更多的法則，並將已經可被正確分類的資料從訓練資料集中移除。為了提升建模效率，本研究將使用F-score 屬性選取方式，來降低資料集的維度，在維持甚至提升正確率的情形下節省計算成本。最後，從UCI機器學習資料庫取得十一個真實世界的資料集，來檢驗本研究提出的方法，並與其他學者提出的分類演算法比較。從實驗結果可以發現，本研究所提出的方法在分類正確率上能擁有良好的表現。 In this study, a classifier called CNFS-OAA has been presented, where modeling procedure is based on complex neuro-fuzzy system (CNFS). The training dataset are divided into multiple binary-class subsets gradually by using one-against-all (OAA), as the training procedure proceeds. The fuzzy rules of CNFS are mined dynamically. In the CNFS modeling procedure, the method of standard particle swarm optimization (SPSO) is used to adjust the premise parameters and the algorithm of recursive least squares estimator (RLSE) is used to adapt the consequent parameters. The method of rules mining for CNFS is that the number of fuzzy IF-THEN rules is incremented dynamically according to the accuracy of classification while training. More rules will be mined when the training accuracy cannot reach the threshold, and the tuples classified correctly will be removed from the training dataset. For the purpose of increasing modeling performance, a method of feature selection called F-score is used to choose useful features and so to reduce the feature dimensions of dataset. By this way, the computational cost can be saved while keeping or even improving the accuracy. In this study, eleven datasets from the UCI machine learning repository have been used to evaluate the approach proposed. The results by the proposed approach are compared with those by other noted approaches. The experimental results show that the approach proposed has fine performance on classification.
Appears in Collections:	[Graduate Institute of Information Management] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	795	View/Open

社群 sharing

Loading...