Abstract Explainable artificial intelligence provides a framework to ensure fairness, reduce bias, and improve machine learning models. However, explainability should not only appear in the model’s outputs but also assist experts in understanding the entire learning process. It enables easier model verification and helps identify relationships learned during training. This study introduces a feature selection approach based on a novel integration of association rule mining and meta-learning, designed to enhance both transparency and performance. We conducted a case study using the UNSW-NB15 dataset combining traditional feature selection techniques with association rule mining to identify additional interpretable features, yielding high predictive performance. The resulting minimal feature subset consists of four features achieving better performance compared to previous studies. Results demonstrate the potential of combining supervised, unsupervised, and interpretable methods to build an explainable meta-learning layer for feature recommendation, contributing to efficient, scalable, and transparent intelligent systems aligned with modern high-performance computing paradigms.

Machine learning is used for detecting malicious activity in cybersecurity contexts since it provides more adaptable models than signature-based solutions. One of the main challenges in applying machine learning to detect malicious activity is the presence of concept drift, which is a change in data distribution over time. Online models that are updated dynamically are usually applied to handle drift. However, these models require new labelled instances to be updated. Reliable labels are typically scarce, expensive to obtain, and not immediately available, which makes building an effective model difficult. In this work, we propose applying online models with conformal prediction , which provides statistical guarantees, to obtain reliable pseudo-labels to update the model and mitigate the absence of ground truth in new data. Although the use of conformal pseudo-labels produces significant improvements in some cases, these are inconsistent across datasets and models, which limits the applicability of the approach.