Interpretable Approach to Predicting Dynamics in Temporal Weighted Networks
A. Kārkliņš (TU Delft - Electrical Engineering, Mathematics and Computer Science)
Huijuan Wang – Mentor (TU Delft - Multimedia Computing)
R.J. Fokkink – Graduation committee member (TU Delft - Applied Probability)
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
Accurately predicting future interactions in temporal networks is essential for diverse applications, including disease modeling and transparent policy-making under regulatory requirements. While much research has focused on topological or temporal link prediction, relatively few studies address temporal weighted link prediction, where both the presence of future links and their intensity are critical. This paper develops an interpretable approach for forecasting weighted links over time, using contact networks that are aggregated into temporal weighted snapshots. By adapting the Self-Driven (SD) and Self-Cross-Driven (SCD) models introduced in Zou et al. (2023), we design prediction methods that incorporate memory decay while preserving explanatory clarity—avoiding the “black-box” nature common to deep learning.
We benchmark four methods - Baseline, SD, SCD, and an extended SCD* - across a range of physical (face-to-face) and virtual (online communication) contact networks, each exhibiting unique structural and dynamic properties when aggregated into temporal weighted networks. Our results show that the SCD model achieves a 32.86% reduction in Mean Squared Error (MSE) over the Baseline, while SD yields a 19.31% improvement. Moving from SD to SCD confers an additional 15.87% decrease in MSE, underscoring the benefits of incorporating both temporal and structural information. Although SCD* introduces further complexity, it did not show any consistent improvements in predictive performance across datasets.
Additional evaluations using the Area Under the Precision-Recall Curve (AUPRC) highlight dataset-specific variability in capturing active links. Correlation analysis reveals that MSE scales with average link weight distributions, whereas AUPRC correlates strongly with the proportion of active links per network snapshot. These findings emphasize that incorporating decay and structural context, in an interpretable manner, significantly enhances predictive accuracy, although parameter tuning remains crucial for different network topologies and interaction patterns.