Neuro-weighted multi-functional nearest-neighbour classification

Background The performance of nearest-neighbour classification is highly sensitive to the quality of data. In order to reduce the impact of the inevitable existence of irrelevant features in data, this paper employs the information learned by neural networks to implement a feature weighting technique and associated weighted multi-functional nearest-neighbour classification method. Method The non-iterative neural networks are easy to implement while enjoying remarkable computational efficiency. In this paper, four non-iterative neural networks (ELM, E-ELM, RAWN and RVFL) are employed to evaluate the significance of features to decisions learned and stored in the parameters of neural networks. Moreover, the bias of the significance of features is comforted by using cross-validation. The resulting feature significance is converted into feature weights to further implement weighted multi-functional nearest-neighbour classification. Result The experimental results demonstrate that the proposed neuro-weighted feature weighting strategy can effectively reduce the impact of irrelevant features and enhance the performance of multi-functional nearest-neighbour classification. Contribution The proposed algorithm explores an avenue to efficiently utilize the learned knowledge of neural networks to reduce the role of irrelevant feature in nearest-neighbour classification.

Automated summary

This paper proposes a method to reduce the impact of irrelevant features in nearest-neighbour classification by utilizing the learned knowledge of neural networks. Experimental results demonstrate that this method effectively reduces the influence of irrelevant features and enhances the performance of nearest-neighbour classification.