Search

As a generalization of K-nearest neighbor K-NN algorithm, the fuzzy K-nearest neighbor fuzzy K-NN algorithm was originally developed by Keller in 1985 to overcome one of the drawbacks of K-NN i.e. all of instances are considered equally important in K-NN. However, fuzzy K-NN algorithm still suffers from the problem of large memory requirement same as K-NN. To deal with this problem, based on fuzzy rough set technique, this paper proposed two condensed fuzzy nearest neighbor methods denoted by CFK-NN1 and CFK-NN2 and a modified fuzzy K-NN. The CFK-NN1 and CFK-NN2 both consists of three steps: 1 obtaining a fuzzy attribute reduct based on fuzzy rough set technique, 2 finding two sets of prototypes, the one is selected from fuzzy positive region corresponding to CFK-NN1 and the other is selected from fuzzy boundary region corresponding to CFK-NN2, 3 extracting fuzzy classification rules with the modified fuzzy K-NN from the two sets of prototypes. Extensive experiments and statistical analysis are conducted to verify the effectiveness of our proposed method. The experimental results and the statistical analysis of the experimental results both demonstrate that the proposed methods outperform other related methods such as CNN, ENN, and ICF et al.

High accuracy and low overhead are two key features of a well-designed classifier for different classification scenarios. In this paper, we propose an improved classifier using a single-hidden layer feedforward neural network (SLFN) trained with extreme learning machine. The novel classifier first utilizes principal component analysis to reduce the feature dimension and then selects the optimal architecture of the SLFN based on a new localized generalization error model in the principal component space. Experimental and statistical results on the NSL-KDD data set demonstrate that the proposed classifier can achieve a significant performance improvement compared with previous classifiers.

Fuzzy rough attribute reduct has been widely used to remove redundant real-valued attributes without discretizing. By now, there are two existing fuzzy rough attribute reduct methods, one is based on dependency function and another based on discernibility matrix. The former proposed by Shen in 2002 can deal with fuzzy decision table FDT with real-valued condition attributes and fuzzy decision attributes. However, this algorithm is not convergent on many real datasets, and the computational complexity of the algorithm increases exponentially with the number of input variables. The latter proposed by Tsang in 2008 can only deal with fuzzy decision table with real-valued condition attributes and symbol-valued decision attributes. In this paper, we extend the latter method and propose two algorithms for calculating all fuzzy rough attribute reducts to deal with fuzzy decision table with real-valued condition and decision attributes. The first algorthim is designed for computing all fuzzy attribute reducts, yet the computation complexity of this algorithm increases exponentially with the number of attributes. The second one which can find one near-optimal reduct is a heuristic variant of the first algorithm. The experimental results show the proposed method is feasible and effective.

Ensemble learning also named ensemble of multiple classifiers is one of the hot topics in machine learning. Ensemble learning can improve not only the accuracy but also the efficiency of the classification system. Constructing the component classifiers in ensemble learning is crucial, because it has direct influence on the performance of the classification system. In the construction of component classifiers, it should be guaranteed that the constructed component classifiers possess certain accuracy and diversity. Based on the confidence degree of classifier, this paper presents an approach consisting of three steps to dynamically integrate rough set reducts. Firstly, multiple reducts are computed. Secondly, multiple component classifiers with certain diversity are trained on the different reducts. Finally, these component classifiers are integrated by adopting dynamic integration strategy. The experimental results show that the proposed algorithm is efficient and feasible.

Feature selection (FE) is a crucial pre-processing in pattern classification. FE addresses the problem of finding the most compact and informative subset of initial feature set to improve the performance of pattern classification system or to reduce the storage requirement. Recently, Yang et al. proposed a wrapper-based feature selection method for multilayer perceptron (MLP) neural networks. The learning speed of the algorithm is very slow, especially for large database, due to iteratively tuning the weight parameters of the networks with back propagation algorithm. In order to deal with this problem, based on extreme learning machine (ELM), we propose a feature selection algorithm which uses a feature ranking criterion to measure the significance of a feature by computing the aggregate difference of the outputs of the probabilistic SLFN with and without the feature. The SLFN is trained with ELM which randomly chooses the weights of hidden layer and analytically determines the weights of output layer. We compared the proposed algorithm with the Yang's work and other three feature selection algorithms. The experimental results show that our proposed method is effective and efficient.

College English Test Band Four (CET4) in China has been a significant impact on evaluating the English preliminary level of a college student or a class. How to improve the college English teaching and go further to raise passing rate of CET4 are a challenge for many colleges and universities. This paper makes an attempt to quantitatively analyze the CET4 and exam-related factors by using fussy decision tree technique and active learning based on uncertainty. Several features are selected to formulate this problem. The weighted margin is proposed as the new uncertainty measure criterion for unlabeled instance, and a density measure is introduced for avoiding selecting isolated instances. Experiments and simulations on different classes of students show the proposed quantitative analysis method is feasible and effective, which can provide teachers with some useful guidelines for how to improve the college English teaching.

There may be many fuzzy attributes in a fuzzy information system. Different fuzzy attribute has different contribution to classification. More important attributes have more contribution than the others to decision-making. In this paper, based on the importance of the fuzzy condition attributes, a new method generating a fuzzy decision tree is proposed, which uses the important degree of the fuzzy condition attribute with respect to the fuzzy decision attributes to select attributes to expand the branches of a fuzzy decision tree. A comparison between the new method and fuzzy ID3 is provided. It is shown that the new method is more efficient than fuzzy ID3.