Fixed​

• All Classes Sets keeps the dialog settings. With All Classes Sets checked, the analyses of each class combination ignored the Algorithm Parameters of the method (they ran with the defaults) and Average replicates. Both are now applied. With ALL, the PLS-DA analyses of each combination now search for the best number of components, as they do without All Classes Sets. Rerun analyses made with All Classes Sets if you want them to use your settings.
• Feature names with feature selection. For Random Forest, XGBoost, BBC and BRF with feature selection on, Top Feature Importance on the analysis card and the Feature importance plot showed the importances under the names of the wrong columns. The numbers were right, the names were not. They now show the selected columns. The card and the Config tab (which now lists the most important features by name instead of Feature 0, Feature 1, …) are corrected for existing analyses too; a plot made before this fix keeps the wrong names, so create it again.

Documentation​

• New pages: Results stage and Results summary, with interactive demos of the analysis card and the summary table.

New​

• Docs link in the app. The top navigation bar has a Docs item that opens this site in a new tab.
• Permutation plot p-values. On the plot card, a p-value of 0.05 or less is now shown in green (the threshold was 0.005). See Reading the p-value.

Changed​

• Dialog settings are applied to the model. These fields were saved but ignored before; they now change the result:

  • LDA: Solver;
  • Logistic regression: Regularization (C), Max Iterations;
  • SVM: Kernel, Regularization (C);
  • Random forest: Number of Trees, Max Depth;
  • XGBoost: Booster, Tree Method.

  The permutation test and the Excel export use the same values.

• SIMCA uses the significance level α. A class model now accepts a sample only inside its acceptance area; a sample accepted by no class is reported as UNKNOWN. Results are stricter than before, so rerun SIMCA analyses you want to compare. See SIMCA.

• XGBoost early stopping no longer looks at the test fold: it uses a validation part of the training rows. Cross-validated metrics are now honest. See Early stopping and validation.

Fixed​

• XGBoost failed on small data sets with early stopping on.
• Balanced Random Forest failed with Max Features = All features, 10 % or 50 %, and with Class Weight = None.
• PLS-DA failed when the component search used F1 Weighted.

Removed from the dialog​

These fields had no effect and are gone: Solver = LSQR (LDA), Outlier Threshold (γ) (PLS-DA), Score Threshold (feature selection) and Sampling Strategy = Minority (Balanced Random Forest, not supported by the library).