Skip to content

Machine learning

Composed builders over the same layers, fed by pure metrics.

Evaluation

BuilderShows
addConfusionMatrixcounts or row-normalised recall, with per-cell labels
addRocCurveROC + chance diagonal, AUC in the legend
addPrCurveprecision–recall + no-skill baseline, AP in the legend
addCalibrationreliability diagram + ECE
addLiftCurvecumulative gain or lift against the fraction targeted
addPredVsActualpredictions against truth with the y=x reference and R²
addResidualsresiduals against the prediction or the sample order
addLearningCurvetrain/validation score vs. training-set size, with CV bands

Explainability & training

addFeatureImportance · addShapBeeswarm · addPartialDependence (PDP + ICE) · addAttentionMap · addTrainingCurves (EMA smoothing + best-epoch marker) · addRidgeline (distributions over epochs) · addEmbedding · addDecisionBoundary

Pure metrics

Classification — confusionMatrix, rocCurve, prCurve, calibrationCurve, classificationReport (per-class precision/recall/f1 plus macro and weighted averages), rocCurveOvR (one-vs-rest with macro/micro AUC), liftCurve, logLoss, brierScore.

Regression — r2, mse, rmse, mae.

Reducers — pca(data, n, d, k), standardize, emaSmooth, beeswarmLayout.

A class the model never predicts scores 0 rather than NaN, so macro averages stay comparable across runs.

MIT licensed. WebGL2 required.