Evolutionary Extreme Learning Machine Based Feature Weighted Nearest Neighbor Classification Algorithm

The traditional rough set theory plays an important role in dealing with imperfect and incomplete information. The traditional fuzzy set theory was proposed as a mathematical tool for dealing with imprecise and uncertain knowledge. Fuzzy-rough sets is a hybridization of fuzzy set theory and rough set theory. Fuzzy-rough sets can flexibly handle vague or imprecise continuous real-valued data. Traditional learning algorithms for neural artificial networks usually needs many parameters than required. However, it is undesirable that the learning algorithm stops at a local minima if it is located far above a global minima. To solve these issues related with gradient-based algorithms, an efficient learning algorithm for feedforward neural networks was proposed, namely Extreme Learning Machine learning algorithms (ELM). The main idea of this is that the input weights and hidden layer biases of ELM can be randomly assigned. Unlike the traditional classic gradient-based learning algorithms, the learning speed of ELM is very fast. Furthermore, the proposed ELM has better generalization performance than the gradient-based learning algorithms. What’s more, the ELM avoids many difficulties such as local minima and overfitting.The quality of features has a significant impact on the performance of a learning algorithm for classification tasks. The classifier performance can be degraded if there are irrelevant features, which are often inevitable in the real application. To solve these issues, a framework of feature weighted is proposed in this paper. What’s more, this framework is used in three different neighbor algorithms. These three different algorithms are as follows:feature weighting using evolutionary extreme learning machine for nearest-neighbor classification, evolutionary extreme learning machine based weighted nearest-neighbor equality classification and evolutionary extreme learning machine based weighted fuzzy-rough nearest-neighbor algorithm. Three algorithms proposed in this paper are used the same framework of feature weighted. The difference of three algorithm is that they are applied to different neighbor algorithms. The K nearest neighbor algorithm is applied in the new framework proposed in first algorithm. The classification performance is enhanced in first algorithm which is proved by experiment. The K nearest neighbor equality algorithm is applied in the new framework proposed in second algorithm. The classification performance is enhanced in second algorithm which is proved by experiment. The fuzzy-rough nearest-neighbor algorithm is applied in the new framework proposed in third algorithm. The third algorithm is proposed to solve the imprecise and uncertain knowledge which is caused by overlapping of training samples and insufficient property in the multi-class problems. The classification performance is enhanced in third algorithm which is proved by experiment.