Characterizing the Feature Space Transformations Produced by Genetic Programming Using the Optimal Transport Dataset Distance
本文用最优传输数据集距离直接量化遗传编程在特征工程中产生的变换幅度，分析其与性能指标在不同分析层次上的相关性，发现群组层次相关性最高。
Genetic programming (GP) has been widely used for machine learning tasks. One of the domains where it has shown the most promise is feature space transformation, also known as feature engineering. Several GP methods have been proposed for this task, all of which assume that GP offers the necessary search capability to find a new representation for the problem data. However, published works have not quantified or characterized the nature of the evolved transformations directly, focusing on an indirect evaluation captured by the empirical predictive accuracy achieved by an additional learning method. The present work addresses this gap, using the recently proposed Optimal Transport Dataset Distance (OTDD) to quantify the magnitude of the evolved transformations. Using M3GP as the method under study, correlation analysis is used to determine how the magnitude of the feature space transformations relate with other performance indicators. The analysis is performed at different levels, including the global level (supervised machine learning), the group level (classification or regression), the problem level and the run level. Results show that the impact of the evolved transformations differs depending on the level of analysis chosen, with the highest correlations observed at the group level.