Comparative analysis of thresholding strategies and the impact of class imbalance on link prediction methods

Authors: Vidza, M.S., Budka, M., Chai, W.K., Thrush, M., Alves, M.T.

Journal: Knowledge Based Systems

Publication Date: 27/09/2026

Volume: 350

ISSN: 0950-7051

DOI: 10.1016/j.knosys.2026.116466

Abstract:

Accurately predicting future links, known as link prediction (LP), is crucial for understanding how temporal networks evolve. This study focuses on the aquaculture sector, where predicting the movement patterns of live fish consignments between fish sites is essential for effective stock management. Given the biosecurity risks associated with inaccurate link prediction, this study evaluates the performance of several LP methods using two primary thresholding strategies: the fixed-threshold and the threshold–curve. Recognising the challenges posed by class imbalance in real-world datasets such as aquaculture networks, the study introduces and evaluates a customised class balancing technique: targeted edge removal. Its impact on LP methods is assessed using evaluation metrics including accuracy, precision, recall, F1 score, and the Matthews Correlation Coefficient (MCC). Among the six LP methods evaluated, the Edge Weighted Katz Index (EWKI) consistently outperforms other methods, achieving a precision of 93.59%, recall of 84.35%, F1 score of 88.73%, and an MCC of 0.89 before class balancing. After applying the class balancing technique, EWKI maintained a robust and stable performance. These findings show how thresholding strategies and class balancing techniques affect LP performance in aquaculture networks, and why they should be considered when applying LP to stock-movement analysis and disease surveillance.

Source: Scopus