Improving representativeness in big data analysis through weighted machine learning methods: A case study on the logistic regression model

Abstract

In the presence of Big Data, it is essential to recognize that despite the abundance of data, these often do not faithfully represent the target populations. Therefore, analyzing these vast datasets does not guarantee representativeness, as they are collected without proper sampling design. Integrating survey weights and auxiliary information into machine learning algorithms constitutes a major challenge in making the samples more representative of the overall population. Moreover, only a few statistical learning software packages offer options to include these weights in their estimation process. In this paper, we introduce a novel weighted configuration of the logistic regression algorithm and employ a bootstrap method to compare its performance against non-weighted models. Our contributions demonstrate the importance and relevance of incorporating different weights for instances and provide a practical approach for analysts in settings where traditional statistical learning tools fall short. This work bridges a critical gap in statistical learning, ensuring that conclusions drawn from large datasets are robust and generalizable.