Volatile fingerprinting and interpretable machine learning for authenticating New Zealand monofloral honeys

Authenticating monofloral honeys is essential for protecting premium markets and ensuring traceability. This study applied an integrated analytical and explainable machine-learning workflow to identify volatile biomarkers for four New Zealand monofloral honeys: thyme, mānuka, kānuka, and clover. Twenty-two samples were profiled using headspace solid-phase microextraction gas chromatography-mass spectrometry (HS-SPME/GC-MS), yielding 122 tentative volatiles across aldehydes, alcohols, acids, esters, terpenoids, and phenolics. Data were analysed using ANOVA, hierarchical clustering, Partial Least Squares Discriminant Analysis (PLS-DA), and Random Forest classification with SHapley Additive exPlanations (SHAP). ANOVA and heatmap analysis revealed honey-specific volatile modules, while PLS-DA confirmed clear supervised separation of floral types. A rule-based SHAP framework was implemented to select biomarkers that were not only influential in the model but also directionally consistent and chemically distinctive. Thyme honey was characterised by short-chain fatty acids and oxygenated terpenoids; mānuka by methoxyacetophenones and benzofuran/methoxylated benzoates; kānuka by anisole-type aromatics and bicyclic monoterpenes; and clover by phenylpropanoid-related aldehydes, fusel alcohols, and linalool-oxide derivatives. These panels achieved non-overlapping group separation and near-perfect cross-validated performance (micro-average ROC-AUC = 0.995). This combined HS-SPME/GC-MS and RF-SHAP approach provides a transparent, statistically supported route to defining interpretable volatile biomarkers, offering a scalable framework for honey authentication, quality assurance, and traceability, and helping safeguard the premium positioning of New Zealand monofloral honeys in global markets.