For each compound, VIP measures how much it contributes to the PLS-DA model's ability to separate classes, combining how strongly the compound loads onto each PLS component with how much variance that component explains. VIP ≥ 1 is the standard rule-of-thumb cutoff (a compound contributing "more than average" across all compounds) — it's a candidate filter, not proof of biological relevance on its own.

For each VIP ≥ 1 candidate, a one-way ANOVA F-test checks whether that compound's area actually differs across your Class groups (not just within the PLS model's internal math). A compound is marked validated: true only if VIP ≥ 1 and p < 0.05 — meaning both the multivariate model and a simple univariate test agree it separates classes. With small sample counts, ANOVA has low statistical power, so a real biomarker can still fail this test on a small dataset — treat the validated list as a shortlist, not a final answer.

SVM (RBF kernel) finds the boundary that maximizes the margin between classes in a transformed feature space — good with clear-margin separation, less interpretable.
Random Forest averages many decision trees, each trained on a random subset of samples/features — robust to noise, gives feature importances "for free".
k-NN (k=5) classifies a sample by majority vote among its 5 nearest neighbours in feature space — simple, but sensitive to feature scaling (we standardize first) and to irrelevant features.
Naive Bayes (Gaussian) assumes each feature is normally distributed within a class and independent of the others, then applies Bayes' rule — fast, works well with small n, but the independence assumption rarely holds exactly for correlated lipid compounds.
ANN (2-layer MLP, 32→16 units) learns a non-linear decision boundary via backpropagation — most flexible, but needs more data to avoid overfitting than the others.

All 5 are evaluated with the same stratified k-fold cross-validation split, so the accuracy/sensitivity/specificity/precision/recall/F1/ROC-AUC columns in the Model versions table are directly comparable across algorithms.

Every pair among the validated biomarkers is tried as a ratio (compound A ÷ compound B), scored by how well that single number separates "contains lard" from "does not" using ROC-AUC (area under the true-positive-rate vs false-positive-rate curve — 1.0 is perfect separation, 0.5 is no better than chance). The pair with the best AUC is kept. Its cutoff is the Youden-optimal point (max of true-positive rate − false-positive rate along the ROC curve) — the threshold that best balances catching real lard against false alarms, not necessarily 50/50.

Every time you click Retrain now, all 5 algorithms are trained on the full current dataset and each gets its own row — so one retrain = 5 new version rows, all sharing the same version_label (a timestamp). That's what the label groups: "which retrain run did this come from." Comparing rows with the same label tells you which algorithm did best on that exact dataset; comparing rows across different labels tells you whether adding more training data made things better or worse over time. Nothing auto-activates — you choose which single version is "live" for predictions.

Accuracy — overall % correct. Misleading if your classes are imbalanced (e.g. mostly halal samples, few lard).
Sensitivity / Recall — of the samples that truly are a given class, how many did the model catch. For a halal-authentication tool, this matters most for the non-halal class: missing real lard (a false negative) is the costlier mistake.
Specificity — of the samples that truly are NOT a given class, how many did the model correctly rule out. High specificity means fewer false alarms on genuinely halal samples.
Precision — of the samples the model called a given class, how many actually were. Low precision means a lot of "suspicious" flags turn out fine on investigation.
F1 — the balance point between precision and recall, useful as a single number when you don't want to prioritize one over the other.
ROC-AUC — how well the model ranks samples by confidence, independent of any specific cutoff; 1.0 is perfect, 0.5 is random guessing.

Given the domain, a reasonable default is: prioritize sensitivity for the non-halal class first (don't miss real lard), then use accuracy/F1 to break ties among versions that are similarly sensitive.

The predicted sample's cleaned compound data gets copied into the training dataset with the Class you confirm (tagged source: promoted so you can tell it apart from directly-added samples). It does not retrain anything automatically — the model stays exactly as it was until you go to Models and click Retrain now again, at which point the promoted sample is included like any other training sample.