Visualizing Pseudo-Boolean Functions: Feature Selection and Regularization for Machine Learning

Abstract

The concept of evolutionary fitness landscapes, along with their visualizations, was introduced in biology by Sewall Wright in the 1930 s. The study of fitness landscapes is also important in artificial intelligence, evolutionary computation, and machine learning. In this paper, we discuss the difficulty of visualizing pseudo-Boolean fitness landscapes in the context of feature selection for machine learning. Visualization techniques for fitness landscapes, specifically hinged bitstring maps and local optima networks, are used to derive information from the landscapes. Specifically, we consider the problem of feature selection for machine learning with random forests in the setting of several real-world datasets as a case study. Using these techniques, we highlight the transformation on the multimodal structure of the feature selection fitness landscapes. This work improves the understanding of visualization for feature selection in machine learning, and promises to lead to improved feature selection and visualization methods.

Visualizing Pseudo-Boolean Functions: Feature Selection and Regularization for Machine Learning