Datasets validation: should we trust publicly available datasets for research?

In this era of data-driven research, we heavily rely on public datasets to treat them as original datasets and develop artificial intelligence (AI) systems. Therefore, it is important to validate the reliability and quality of public datasets to build trusted machine learning (ML) models using them. The understanding of the compressed domains of these original datasets can help us detect the features that contribute to their reliability and quality. Thus, this knowledge can support the development of compressed ML models to reconstruct more robust, reliable and high-quality distilled versions of the original datasets. Hence, we can adopt the concept of compressed sensing and compressed learning to develop a computational framework that can validate publicly available datasets, before using them to build trusted AI systems. In this paper, an innovative computational framework that combines compressed sensing and compressed learning with kernel density estimation (KDE) technique is proposed. This computational framework can identify and discard unreliable and redundant features, as well as noise, and reconstruct robustly distilled versions of the original data by retaining their core statistical characteristics. This KDE-based computational framework supports both data validation and AI explainability. The KDE technique allows the productive use of its bandwidth to generate explainability through transparent feature contributions. In our simulations, we have used Kernel Ridge Regression and Deep Learning models for compressed sensing, and Random Forest classifiers for compressed learning. These models effectively validate and evaluate the proposed computational framework. Additionally, Sparse Random Projection is used to construct compressed datasets, while the Mean Squared Error technique evaluates whether the distributions of the reconstructed datasets align with that of the original datasets. Simulations using public datasets confirm that the proposed computational framework effectively delivers compressed sensing models capable of quantifying the reliability and quality of these datasets; hence, these models advance data-driven research.