Self-Organizing Transformations for Automatic Feature Engineering

Silva Rodrigues E, Martins DML, Lima Neto FB

Abstract

Feature Engineering (FE) consists of generating new, better features to improve Machine Learning models. Very often, FE is performed in a series of trial-and-error steps conducted manually by data scientists. Moreover, FE requires data-specific and domain knowledge, both rarely easy to acquire. To alleviate these problems, we propose the Self-Organizing Automatic Feature Engineering (SOAFE), a novel approach for Automatic Feature Engineering (AFE). Different from the majority of the existing AFEs, SOAFE employs an unsupervised technique (Self-Organizing Maps) to identify patterns in the data, and apply a form of cooperative training, inspired by Generative Adversarial Networks, to improve the feature construction. Our results on several datasets show that SOAFE can improve classification models when compared with existing AFE approaches.