Key points are not available for this paper at this time.
This study investigates whether network centrality metrics enhance machine learning-based risk classification in international food safety notification systems. We constructed a directed, weighted country network from 1,364 notifications spanning 96 countries (2020–2025) and computed 11 centrality measures, including PageRank, betweenness, and eigenvector centrality. Although centrality features are strongly associated with risk categories (Kruskal-Wallis p<10−16, η2=0.067), they provide negligible predictive contribution (ablation ΔF1=−0.005). This mismatch is explained by redundancy analysis: categorical country encodings correlate highly with network position (ρ=0.87 for Spain’s out-strength), rendering explicit centrality metrics largely superfluous for tree-based classifiers. The network shows scale-free tendencies, moderate community structure (modularity Q=0.293), and negative assortativity (r=−0.26). Gradient Boosting achieved F1-macro =0.459 (0.492 with SMOTE), Cohen’s κ=0.538, and ROC-AUC =0.852; year dominated feature importance (0.079), indicating regulatory evolution rather than topology as the primary signal. Temporal validation produced a 16.6 % F1 degradation, with a disproportionate minority-class decline (−31%). Hierarchical classification improved minority detection (no-risk recall: 0 % → 25 %). Community-stratified evaluation uncovered fairness disparities (F1 range =0.26) not explained by sample size (ρ=0.31, p=.61). Overall, statistical significance does not imply predictive importance for network features in food safety classification. Network analysis remains valuable for interpretability and identifying structurally significant nodes, while comparable predictive performance can be achieved using simpler categorical encodings. Severe class imbalance (77:1) and concept drift motivate hierarchical modeling, quarterly retraining, and human oversight for minority classes.
Shinyclimensa et al. (Thu,) studied this question.