Results summary
The Analysis Performance Summary at the top of the Results stage puts the cross-validated accuracy of every classification analysis of the experiment into one table. Use it to compare methods on the same data and to see how the accuracy changes when you add or remove classes.
Analysis Performance Summary
4 label configurations, 7 methods| Label Column | Included Classes | Data Range | Description | KNN | LDA | LOGISTIC | PLSDA | RANDOM_FOREST | SIMCA | SVM | Actions |
|---|---|---|---|---|---|---|---|---|---|---|---|
| class_Origin | All classes: China, India, Kenya | D1:G48 | Three origins, RGB + intensity | ||||||||
| class_Origin | China, India | D1:G36 | Click edit to add description... | ||||||||
| class_Origin | China, Kenya | D1:G21, D37:G48 | Click edit to add description... | ||||||||
| class_Origin | India, Kenya | D22:G48 | Click edit to add description... |
Summary table
What is included. Every completed classification analysis: LDA, Logistic, SVM, Random Forest, KNN, PLS-DA, SIMCA, XGBoost, BBC and BRF. PCA is not included (it has no accuracy). An analysis is added when it completes and removed when it fails or is deleted. With no classification analyses, the card says No classification analyses found.
Rows. One row per configuration: the same Label Column, the same Data Range and the same Included Classes. Rows are sorted by label column, then by the number of included classes, most classes first.
| Column | Content |
|---|---|
| Label Column | The label column of the analyses. |
| Included Classes | The classes used. All classes: … when every class of the label column is included. Hover to see the included and the excluded classes. |
| Data Range | The data range, cut to 40 characters. Hover to see all of it. |
| Description | Your note for the row. Click the pencil, type, and click Save. The description is saved on the server and goes into the Excel export. |
| one column per method | LDA, PLSDA, RANDOM_FOREST, … in alphabetical order. Only methods that were run appear. |
| Actions | Delete All with the number of analyses in the row. See Deleting a configuration. |
Cells. Each cell lists every run of that method in that configuration, oldest first. One entry is:
- ✓ — the run completed;
- the CV accuracy in percent: the accuracy of the test folds, averaged
over the folds (
cv_mean, card label CV Accuracy; see Metrics); ±the standard deviation over the folds. It is hidden when it is 0.
Click an entry to scroll to its analysis card. A dash (—) means the method was not run in this configuration.
The colour of the accuracy:
| Colour | CV accuracy |
|---|---|
| Red | above 99 % |
| Green | 90 % – 99 % |
| Blue | 80 % – 90 % |
| Yellow | below 80 % |
Red does not mean "bad": it marks a perfect or nearly perfect result, which is worth checking. See Comparing models.
The header shows how many configurations and methods the table has, and how many runs it contains (N completed). Click the header to fold or unfold the table.
- A single analysis adds an entry to the row of its label column, data range and checked classes.
- ALL adds one entry per method (PCA excluded) to one row.
- All Classes Sets adds one row per combination of classes. Each row has its own data range, cut down to the rows of its classes. See Running several methods at once.
A row is matched by the exact data range text and the exact list of
classes. If what looks like the same configuration appears as two rows, the
range is written differently (for example D1:G36 and D1:G20, D21:G36) or
the classes are listed in a different order.
Deleting a configuration
Delete All in the Actions column deletes every analysis of the row. A confirmation, Delete Configuration Analyses, shows the label column, data range, classes, description and the number of analyses. Their plots, predictions and saved models are deleted as well, and the row disappears. This cannot be undone.
Comparing models
The table shows only the mean CV accuracy. Before you pick a model, keep in mind what it does not show.
- Compare within a row. Runs in one row used the same rows and classes, so their accuracies are comparable. Between rows the task is different: with two classes, random guessing gives about 50 %, with three about 33 % (for equal class sizes).
- Look at the
±. It is the spread over the folds. If two methods differ by less than their standard deviations, the data do not show that one is better. - The row does not record everything. Two runs of the same method land in the same cell even if they used different CV schemes, numbers of folds, method parameters, feature selection or replicate averaging. Check the Config tab of each card, and use the Description to note what you changed.
- Accuracy hides class balance. With classes of different sizes, a model can reach a high accuracy by favouring the large class. Check sensitivity, specificity and the confusion matrix on the card. See Metrics.
- Check red (above 99 %) results. A perfect CV accuracy on small data often means that the test folds were not independent, for example replicates of one sample in both training and test folds. Check the CV scheme (Group K-Fold and replicates), run a permutation test, and read Cross-validation and data leakage.
Export
Export on the Data stage writes the summary to the Analyses_Summary sheet of the Excel workbook. The sheet differs from the table on the page:
- every run has its own row (a row per configuration and run number), so a cell holds one value;
- accuracies are numbers from 0 to 1, not percentages, and there is no standard deviation;
- the columns are Label Column, Data Range, Included Classes, Description, then one column per method.
The metrics of every analysis are on the Analyses_Details sheet. See Exporting analysis summaries.