A Tabu search heuristic for smoke term curation in safety defect discovery

Abstract

The ability to detect and rapidly respond to the presence of safety defects is vital to firms and to regulatory agencies. In this paper, we employ a text mining methodology to generate industry-specific “smoke terms” for identifying these defects in the countertop appliances and over-the-counter medicine industries. Building upon prior work, we propose several methodological improvements to enhance the precision of our industry-specific terms. First, we replace the subjective manual curation of these terms with an automated Tabu search algorithm, which provides a statistically significant improvement over a sample of human-curated lists. Contrary to the assumptions of prior work, we find that shorter, targeted smoke term lists produce superior precision. Second, we incorporate non-textual review features to enhance the performance of these smoke term lists. In total, we find greater than a twofold improvement over typical human-curated lists. As safety surveillance is vital across industries, our method has great potential to assist firms and regulatory agencies in identifying and responding quickly to safety defects.