Improving Automated Feature Engineering Using Meta-Learning Based Techniques

Abstract

Building well-performing machine learning pipelines requires the use of feature engineering. However, building highly predictive features takes time and requires subject-matter expertise. Although automated feature engineering research has recently gained a lot of attention from both academia and industry, the scalability and efficiency of the current methods and tools are still essentially subpar. To this end, we proposed meta-learning techniques to improve the performance of two automated machine learning frameworks; BigFeat and AutoFeat. Extensive experiments were conducted on 17 and 10 datasets for Bigfeat and AutoFeat, respectively. The results show that the proposed meta-learning techniques achieved an average improvement of F1-Score = 1.51% on BigFeat and an average improvement of F1-Score = 1.11% on AutoFeat.