Hybridization of feature selection and feature weighting for high dimensional data

摘要

The classification of high dimensional data is a challenging problem due to the presence of redundant and irrelevant features in a higher amount. These unwanted features degrade accuracy and increase the computational complexity of machine learning algorithms. In this paper, we propose a hybrid method that integrates the complementary strengths of feature selection and feature weighting approaches for improving the classification of high dimensional data on the Nearest Neighbor classifier. Specifically, we suggest four strategies that combine filter and wrapper methods of feature selection and feature weighting. Experiments are performed on 12 high dimensional datasets and outcomes are supported by Friedman as well as Holm statistical tests for validation. Extended Adjusted Ratio of Ratios is used to recognize the best method considering accuracy, feature selection, and runtime. The results show that two proposed strategies outperform other well-known methods in accuracy and features reduction. The hybrid feature selection-feature weighting wrapper method is the best among all in accuracy while the hybrid feature selection filter-feature weighting wrapper method is the most suitable for reducing features and runtime. Thus, the promising outcomes validate the importance of hybridizing feature selection and feature weighting while dealing with high dimensional data.