Reproducible and Explainable Active Learning for Systematic Reviews: An Applied Data Science Contribution Uniting Academic and Industrial Methodologies

Abstract

Systematic reviews are the gold standard for synthesizing scientific evidence, yet the explosion of published research is turning their manual screening phase into an unsustainable bottleneck. Data Science offers a solution: Active Learning. This machine learning technique incrementally learns from the user to prioritize relevant records, without needing large prelabeled datasets. Tools like ASReview introduce this technology, but its implementation is far from complete, and deploying it effectively in academia requires more than raw algorithmic performance. This book details an academic Applied Data Science project that improves the effectiveness of Active Learning within ASReview, following the technology as its implementation matures. While industry projects often focus on performance metrics, the academic setting introduces additional demands for explainability and reproducibility. Simultaneously, the applied nature of the work requires delivering actionable advice. To define how to execute an applied study in this environment, this dissertation engages with four core themes: Human-Centered Design, Software Usability, Reproducibility and Evidence, and FAIR data.