xAutoGA: an automated deep learning software to educate users

Automated deep learning (AutoDL) has emerged as a promising approach to simplify model development for non-experts. However, existing automated tools often lack critical domain-specific features, such as automatic sliding window optimization and subject-wise data splitting, and they remain opaque, which limits their adoption in scientific applications. This study introduces xAutoGA (for E x plainable Auto DL based on a G enetic A lgorithm), designed to address these limitations for multivariate time-series data. The software automates subject assignment to the training, validation, and test sets and jointly optimizes windowing hyperparameters and model architectures in both time and frequency domains. To address the black-box issue, the software guides users through an interactive interface that allows them to configure hyperparameters for the genetic algorithm and neural networks. In addition, it employs a natural selection analogy to explain the selection of specific hyperparameters by visualizing the optimization trajectory. To further enhance interpretability, a rule-extraction method identifies the hyperparameters contributing to top-performing models, which are then synthesized into educational insights using a Large Language Model (LLM). The proposed software is benchmarked against two existing AutoDL systems (McFly and DRAGON) for activity recognition and movement disorder detection tasks. Feedback on the educational component and software usability was collected from 19 university students and two machine learning experts. Statistical tests demonstrated superior performance on three of the five public datasets compared to existing tools ( p -value< 0.05), and user feedback supported the educational value of the proposed software.