Skip to main content

Analysis dialog

The analysis dialog turns the data table into a configured, running analysis. It has the same shape for every method: an Analysis Name field, then four collapsible sections — Data Selection, Preprocessing, Algorithm Parameters (skipped for the batch button, see below) and Cross-Validation — and Cancel / Start Analysis at the bottom.

Opening the dialog​

Click a method button in Analysis Tools, on the left of the Data stage (PCA, LDA, Logistic, SVM, RF, KNN, PLS, SIM, XGB, BBC, BRF, or ALL — see Running several methods at once). The dialog title becomes "Configure Method Analysis" and the Analysis Name defaults to "Method Analysis"; both are plain text, so you can rename the analysis before you start it.

Clicking a method button also saves the table to the server first, the same save that happens on Save Changes — see Edits are kept in the browser until they are saved. The analysis always reads the saved table, never unsaved edits.

If you close the dialog and open it again for another method in the same visit to the page, it remembers your last Cross-Validation settings and the All Classes Sets checkbox. The data range, label column, feature columns and algorithm parameters are always recomputed for the new method.

Data range and label column​

Data Range (Excel notation, e.g. D1:G9) and Label Column are the two required fields. By default the range covers every data column and every row of the table, and the label column is the first classification column, or sample_name if the table has none. Click the cursor button to pick a range with the mouse instead of typing it — see Selecting a data range for that flow — or click Preview to check what the current range resolves to (row and column counts, number of data channels, and a warning if it spans several disjoint blocks).

Label Column offers every sample-info and classification column of the table: sample_name, cell_coordinate, and one entry per class type (labeled "Type (Classification)"). cell_coordinate is listed but not a meaningful label — every well has a different one, so no method can learn anything from it.

Once a label column is chosen, a Label Statistics panel and a Filter by Column Labels checklist appear (Column is the internal column name, for example class_Origin). The checklist decides which rows enter the analysis: every value of the label column found in the current range gets a checkbox and a count, and unchecking one removes its rows from the data sent to the backend (it edits Data Range for you, splitting it around the excluded rows). At least one label must stay checked.

Column containing sample names or class labels for analysis
Uncheck labels to exclude their rows from analysis
6
3
Label Statistics:
2 unique labels9 total samples
Label distribution: China: 6, India: 3
0 row(s) are excluded from the analysis while it stays unchecked.
Interactive demo: pick which Origin labels enter the analysis.

Changing Label Column resets this checklist to "all included" for the new column, and also resets CV Method and Number of Folds to their defaults for that column — Group K-Fold with one fold per unique sample name, or Random K-Fold with folds = rows ÷ 6 for sample_name. Both are yours to change afterwards. The full rule, and why Group K-Fold is the default, is in Cross-validation.

The All Classes Sets checkbox next to the checklist disables the individual checkboxes (row filtering by hand and batch class enumeration are mutually exclusive) — see Running several methods at once.

note

Label Statistics flags missing values and warns if the column has fewer than two unique values in the current selection — classification needs at least two classes, and PCA needs the label only to colour the score plot.

Selecting feature columns​

Between the label column and the label filter sits a second checklist, Feature Columns. It is folded by default: click its title (with the chevron) to open it. It has one checkbox per data column of the original range (the range you had before any filtering). Unchecking a column removes it from Data Range the same way unchecking a label removes rows — the two checklists write to the same range string independently, so filtering columns never undoes a row filter and vice versa. Typing a range into Data Range by hand instead overrides both checklists' contribution until you touch a checkbox again.

Uncheck columns to exclude them from the analysis
Interactive demo: click Feature Columns to fold the list, uncheck a column to exclude it from the analysis.

At least one feature column must stay checked. If you deselect columns scattered across the range, the resulting comma-separated range string can get long; past 500 characters the dialog asks you to deselect fewer columns or simplify the label filter instead.

Not the same as feature selection

This checklist removes columns before anything is sent to the backend. Enable Feature Selection, in Preprocessing below, is a statistical selection (kk best of whatever columns remain) applied after — see Feature selection. Uncheck a column here only when you are sure it should never be considered; use feature selection when you want the data to decide.

Method parameters​

Preprocessing applies to every method, including the batch button:

  • Average technical replicates groups rows with the same sample_name and averages them before the model sees the data — see Averaging replicates. It is disabled when the label column is sample_name and the method is LDA, which needs at least two samples per class.
  • Enable Feature Selection turns on ANOVA F-test or mutual-information scoring, with Selection Method and Number of Features — see Feature selection.

Algorithm Parameters holds the fields specific to the chosen method: number of components, kernel, tree depth, and so on, plus Cross-Validation (CV Method, Number of Folds) below it for every method except PCA. The full parameter table, its defaults and its limits are on that method's own page (PCA, LDA, Logistic Regression, SVM, Random Forest, KNN, PLS-DA, SIMCA, XGBoost, BBC, BRF — see Chemometrics methods). As one example, PLS-DA's fields:

Will be determined automatically via search.
Automatically find the best number of components via cross-validation.
Interactive demo: PLS-DA's Algorithm Parameters. Other methods show different fields, listed on their own page.

A few numeric fields track the width of Data Range automatically: PCA, PLS-DA and SIMCA's number of components is capped at the number of feature columns, and PLS-DA/SIMCA start at min⁡(p,7)\min(p, 7). Changing the range keeps a value you have edited by hand, but re-derives one you have not touched yet. Every field is validated on Start Analysis (and by Preview for the data range and label column); an invalid value is reported under the field itself, and the dialog will not submit until it is fixed.

Running several methods at once​

The ALL button, next to the "Analysis Tools" header on the Data table, opens the same dialog for All Types. Algorithm Parameters is hidden — there is nothing to configure per method — but Data Selection, Preprocessing and Cross-Validation are shared by all of them. One click creates eight analyses with that shared configuration: PCA, LDA, Logistic Regression, SVM, Random Forest, KNN, SIMCA and PLS-DA, named "Name - TYPE". XGBoost, BBC and BRF are not part of the batch and must be run individually. The PLS-DA analysis in the batch always searches for the optimal number of components (accuracy, 2 to 10); PCA in the batch ignores the Cross-Validation settings, same as it does on its own.

All Classes Sets, next to the label checklist (above), runs the chosen method — or all eight, if combined with ALL — once per combination of two or more of the label column's unique values anywhere in the experiment (not just the rows currently selected): every pair, every triple, and so on up to the full set. Each analysis is named "Name - TYPE - label1+label2" (or "Nclasses" for four or more) and gets its own row filter and its own number of CV folds, recalculated for just its rows. Every analysis otherwise uses the settings of the dialog: algorithm parameters, preprocessing and cross-validation scheme.

6
3
4
3 classes → 4 analyses per method, one per combination of 2 or more classes:
China + IndiaChina + KenyaIndia + KenyaChina + India + Kenya
With All Classes Sets checked, the checklist is read-only: every combination gets its own analysis.

With nn classes this creates 2n−n−12^n - n - 1 analyses per method — 4 for three classes, 11 for four, 26 for five. Check the count before you click Start Analysis with many classes: it grows quickly, and combined with ALL it multiplies by up to eight methods.

Analysis status​

Start Analysis validates the form, closes the dialog, and queues one or more Analysis rows with status pending (or queued, for a batch created by ALL or All Classes Sets). They run in the background and move through pending → running → completed or failed; open the Results stage to watch them and read the output. A failed analysis keeps the table and every other analysis untouched — only that one row gets an error message.

note

While a batch is queued, its rows briefly show status queued with a message like "Batch 2/5: Starting" in the field that normally holds the error — that is progress information, not a failure, and it is replaced once the analysis actually starts running.