Research And Application On Rough Set Based Feature Selection Algorithm

Feature selection algorithm can reduce feature dimension by eliminatingirrelevant or redundant features. This is essential in machine learning and patternrecognition. Because a good feature selection algorithm helps a lot in constructing acomprehensive learning model with good generalization ability. Two main points ofthis paper: i) compare two kinds of supervised feature selection algorithms; ii) studysemi-supervised feature selection algorithm with Fuzzy Rough Set theory.Recently, researchers have done a lot of work on supervised feature selectionalgorithms, among which Margin based and Fuzzy Dependency based methods arebrilliant. But till now, no work has been done to compare these two kinds of methodsin detail.We focus on4different algorithms, Margin based-Simba and Relief; and FuzzyDependency based-WDL-MFD，FD-Ranking. We analyze and compare them withthe following aspects: search strategy used, feature numbers selected, classificationaccuracy on classifiers, ability to eliminate random noise features and computationalcomplexity. Experiments show that all the algorithms can select features with highdiscriminatory power.We introduce Fuzzy Rough Set theory to semi-supervised feature selection.Semi-supervised learning cries for new feature selection algorithms, takingcomprehensive utilization of both class information provided by labeled data and datastructure presented by unlabeled data. In this paper, we combine Fuzzy Rough Settheory with Spectrum Analysis theory, and try to propose an improvedsemi-supervised feature selection algorithm. At last, we apply this method to choosebest features to classify two kinds of stars.