Graph-enhanced machine learning for predicting food safety risk severity from multi-layer RASFF notification networks

Food safety surveillance systems generate large volumes of notification data that encode relationships between countries, hazards, and food products, yet predictive models rarely use this relational structure. This study proposes a network-feature-enhanced machine learning framework for the European Rapid Alert System for Food and Feed (RASFF), in which descriptors derived from notification networks are supplied as input features to conventional classifiers. Using 28,708 notifications (2019–2025) whose free-text origin field was harmonised against the ISO 3166–1 country list (626 raw strings mapped to 163 countries), we construct three network layers: a country-to-country notification network, a same-notification hazard co-occurrence network, and a product-category–origin bipartite network. The country network (163 nodes, 1,663 edges) is highly clustered (average clustering 0.714; average path length 2.01), small-world ( \(\sigma = 2.38\) against a degree-preserving null) and modular (nine Louvain communities, \(Q = 0.310\) ). Node centralities (weighted degree, betweenness, closeness, PageRank, eigenvector, HITS hub/authority, clustering) and community co-membership are used as features for five classifiers—Logistic Regression, Random Forest, Gradient Boosting, Extra Trees, and a Multi-Layer Perceptron—to predict the binary regulatory risk decision (serious vs. not serious) for 25,615 notifications. All network features, including Louvain communities, are rebuilt within each training fold of a 5-fold stratified cross-validation to prevent information leakage, and a regulatory classification field that nearly determines the outcome is excluded. Country-network features improved macro F1 for all five classifiers; the best model, Gradient Boosting, rose from 0.704 to 0.763 (AUC 0.816 to 0.867; paired percentile-bootstrap 95% CI for the gain \([+0.052,+0.064]\) , 2,000 resamples), and hazard- and product-layer features added a further 0.014 (95% CI \([+0.010,+0.017]\) ). These findings demonstrate the potential utility of network-derived features for prioritising RASFF notifications and provide a framework for RASFF-based food safety surveillance; validation on other surveillance systems is needed before broader generalisation.

Authors

Institutions

Publication Details

Journal
Discover Food
Published
2026-11-13
DOI
https://doi.org/10.1007/s44187-026-01308-x
Primary Topic
Food Safety and Hygiene
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Graph-enhanced machine learning for predicting food safety risk severity from multi-layer RASFF notification networks

Ramireddy Suneel, K. Lilly Grace Mary, V. Suryanarayana Reddy, T. Vinod Kumar et al.
6 links
Discover Food
Food Safety and Hygiene
article

Graph-enhanced machine learning for predicting food safety risk severity from multi-layer RASFF notification networks

Ramireddy Suneel, K. Lilly Grace Mary, V. Suryanarayana Reddy, T. Vinod Kumar, M. V. Prabath Kumar, A. Yadunanda Kumar
article en

Abstract

Food safety surveillance systems generate large volumes of notification data that encode relationships between countries, hazards, and food products, yet predictive models rarely use this relational structure. This study proposes a network-feature-enhanced machine learning framework for the European Rapid Alert System for Food and Feed (RASFF), in which descriptors derived from notification networks are supplied as input features to conventional classifiers. Using 28,708 notifications (2019–2025) whose free-text origin field was harmonised against the ISO 3166–1 country list (626 raw strings mapped to 163 countries), we construct three network layers: a country-to-country notification network, a same-notification hazard co-occurrence network, and a product-category–origin bipartite network. The country network (163 nodes, 1,663 edges) is highly clustered (average clustering 0.714; average path length 2.01), small-world ( \(\sigma = 2.38\) against a degree-preserving null) and modular (nine Louvain communities, \(Q = 0.310\) ). Node centralities (weighted degree, betweenness, closeness, PageRank, eigenvector, HITS hub/authority, clustering) and community co-membership are used as features for five classifiers—Logistic Regression, Random Forest, Gradient Boosting, Extra Trees, and a Multi-Layer Perceptron—to predict the binary regulatory risk decision (serious vs. not serious) for 25,615 notifications. All network features, including Louvain communities, are rebuilt within each training fold of a 5-fold stratified cross-validation to prevent information leakage, and a regulatory classification field that nearly determines the outcome is excluded. Country-network features improved macro F1 for all five classifiers; the best model, Gradient Boosting, rose from 0.704 to 0.763 (AUC 0.816 to 0.867; paired percentile-bootstrap 95% CI for the gain \([+0.052,+0.064]\) , 2,000 resamples), and hazard- and product-layer features added a further 0.014 (95% CI \([+0.010,+0.017]\) ). These findings demonstrate the potential utility of network-derived features for prioritising RASFF notifications and provide a framework for RASFF-based food safety surveillance; validation on other surveillance systems is needed before broader generalisation.

Discover Food
Government of Andhra Pradesh (IN)
Good health and well-being, Zero hunger
Food Safety and Hygiene
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon