| 摘要: | 資訊科技的蓬勃發展導致資料快速地增加，而這些資料中可能隱含有價值的資訊，機器學習與資料探勘是能從資料中萃取知識與法則的工具。而分類是其中很重要的議題，透過建構好的分類器將資料正確地歸類是很重要的。然而過去許多著名的分類器如支援向量機、類神經網路與K個最臨近點分法有很好的分類結果，但缺乏以人能理解的方式呈現。因此本研究提出融入後設認知策略的複數模糊認知圖 (Metacognitive complex fuzzy cognitive map, McCFCM)，其結合後設認知 (Metacognitive)、模糊認知圖 (Fuzzy cognitive map, FCM)與複數型模糊集合 (Complex fuzzy sets)，並應用於分類問題中。在MCCFCM建模上，分成參數學習與結構學習。參數學習中，使用標準粒子群最佳化演算法 (Standard particle swarm optimization, SPSO)來調整複數型模糊集合位置與FCM的連結權重；結構學習中，使用二元粒子群演算法 (Binary particle swarm optimization, BPSO)調整FCM連結架構。為了建出更強健的器分類器，本研究使用一對多 (One-against-all, OAA)的訓練方法，同時使用費雪分值 (Fisher score, F-score)挑選出重要的屬性。本研究使用加州大學爾灣分校 (University of California-Irvine)的機器學習資料庫中10個資料集來驗證本研究提出之方法，並與其他著名的研究方法比較分類結果。
The rapid development of information system has led to increase a large number of data, and these data usually imply valuable information, for which machine learning and data mining are useful tools to extract knowledge and rules from data. Classification is one of important issues, and it is important to construct a good classifier that can classify the data correctly. Although some famous classifiers had been presented, such as support vector machine (SVM), artificial neuro network (ANN), and K-nearest neighbors (KNN), they lack of being understood by people. Therefore, in this study, to classification problems, we propose a metacognitive complex fuzzy cognitive map (McCFCM) that combines metacognitive, complex fuzzy sets and fuzzy cognitive map. The modeling of McCFCM classifier comprises the phases of parameter learning and structure learning. In the parameter learning phase, the method of standard particle swarm optimization (SPSO) is use to adjust the location of the complex fuzzy sets and the weights of all the connections in FCM
In the structure learning phase, the algorithm of binary particle swarm optimization (BPSO) is used to establish or erase some connections in FCM. In order to build a more robust classifier, we also use one-against-all (OAA) to decompose the dataset whose data are with multiple classes into several binary-class subsets, and the Fisher score (F-score) is used to pick important features for the problem of classification. In this study, ten datasets from the University of California-Irvine (UCI) machine learning repository have been used to evaluate the performance by the proposed McCFCM classifier, whose results are compared with those by other noted classification algorithms. |
