Logo image
Applying Electromagnetism-like Mechanism Algorithm to Feature Selection: Theory and Application
Dissertation

Applying Electromagnetism-like Mechanism Algorithm to Feature Selection: Theory and Application

Lin, Hung-Chun
Doctor of Philosophy (PHD), 國立清華大學, 工業工程與工程管理學系
2010

Abstract

屬性篩選 NP完備問題 電磁演算法 隨機性缺失 穩健性貝氏分類器 Feature selection NP-complete problem Electromagnetism-like Mechanism algorithm Missing at random Robust Bayes Classifier
Nowadays, high dimensional problems have been increasingly occurring in field of data mining, which increases the computation time and cost. In order to deal with this kind of problems, various feature selection methods have been developed in recent years. Based on the differences between computations, the feature selection techniques could be grouped into two: the filter model and the wrapper model. Among them, the wrapper model regards the error produced by the classifier as a criterion for feature selection and focuses on minimizing the miss-classification of the classifier. The issues about features selection are regarded as NP-complete problems. Therefore, many meta-heuristics are proposed for feature selection. The Electromagnetism-like Mechanism (EM) algorithm is proposed by Birbil and Fang in 2003. It makes use of the attraction-repulsion mechanism of the electromagnetism theory to find the optimal solution. So far, EM has been applied to optimization in continuous space and discrete problems, yet the study on feature selection is not found. This study applies EM and combines 1-nearest-neighbor (1NN) for feature selection and classification. A numerical experiment is carried out to verify the feasibility of the EM algorithm with complete data. Then, a real case concerning gestational diabetes mellitus is introduced, and the outcomes demonstrate that the proposed method is workable in the real world case. On the other hand, actual data sets are generally incomplete because of various reasons. Consequently, algorithms for classification issues with incomplete data have received increasing attention. Many methods have been developed to deal with incomplete data. However, these approaches have either some drawbacks or the pre-assumption of missing at random (MAR) for the data, which is difficult to verify. Ramoni and Sebastiani presented Robust Bayes Classifier (RBC) which could eliminate the assumption. Nevertheless, RBC assumes that the attributes are independent for each class. If this assumption is violated, the performance of classification would be degenerated. Therefore, this study applies the combination of the EM algorithm and RBC to find the feature subset with the best performance. Another numerical experiment is carried out to verify the feasibility of EM for feature selection with incomplete data. The implementation results of above two issues showed that the EM algorithm is useful and effective for feature selection with both complete and incomplete data.

Metrics

1 Record Views

Details

Logo image