Archive-based multiple feature construction method using adaptive Genetic Programming
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
The quality of features is an important factor that affects the classification performance of machine learning algorithms. Feature construction based on Genetic Programming (GP) can automatically create more discriminative features, sometimes greatly improving classification performance. However, constructing a single feature or a small number of features may make the linkage information between labels and features insufficient, resulting in poor classification performance, so we introduce a multi-feature construction method. In addition, premature convergence of the GP may also affect classification performance. This paper proposes an archive-based multiple feature construction method which uses elite archive strategy to preserve and select effective constructed features, and employs an adaptive strategy for GP to adjust the crossover and mutation probabilities based on fitness values. Experiments on ten datasets show that our proposed archive-based multiple feature construction method without using adaptive GP can significantly improve the classification performance compared with traditional single feature construction method, and the classification performance can be maintained or further improved by adding the adaptive strategy. The comparisons with three state-of-the-art techniques show that our proposed method can significantly achieve better classification performance.