Improving Generalization of Evolutionary Feature Construction with Minimal Complexity Knee Points in Regression

Genetic programming-based evolutionary feature construction is a widely used technique for automatically enhancing the performance of a regression algorithm. While it has achieved great success, a challenging problem in feature construction is the issue of overfitting, which has led to the development of many multi-objective methods to control overfitting. However, for multi-objective methods, a key issue is how to select the final model from the front with different trade-offs. To address this challenge, in this paper, we propose a novel minimal complexity knee point selection strategy in evolutionary multi-objective feature construction for regression to select the final model for making predictions. Experimental results on 58 datasets demonstrate the effectiveness and competitiveness of this strategy when compared to eight existing methods. Furthermore, an ensemble of the proposed strategy and existing model selection strategies achieves the best performance and outperforms four popular machine learning algorithms.