The advent of big data has catalyzed the development of numerous data-driven applications across various fields. Essential to the efficacy of these applications is the ability to process and analyze high-dimensional datasets, such as time series, images, and beyond. This research focuses on advancing the methodologies for feature extraction and selection tailored to the unique challenges presented by such data. While feature extraction and selection have been considerably investigated for tabular data, their application to high-dimensional data still needs to be examined. In this paper, we provide a protocol to both reduce dimensionality and retain critical information from high-dimensional datasets necessary for modeling and prediction tasks—specifically designed for, although not limited to, symbolic approaches—using, at a lower level, state-of-the-art feature selection algorithms for tabular data in a systematic way. We demonstrate the effectiveness of our approaches using real-world examples and evaluate their performance using appropriate, objective metrics, and we provide a complete open-source package as part of a long-term project for symbolic non-tabular data representation and learning. Our results contribute to the ongoing discourse on data science and machine learning practices, with particular attention to symbolic approaches.
Cavina et al. (Fri,) studied this question.