Machine Learning

The Machine Learning panel turns bags into a classifier. You pick two or more bags, choose which features the model should look at, and train a selector: a named model that can then be pointed at other bags, or at the whole map, from the Selector Panel.

It is a supervised workflow built on your own labeling. If you have not made bags, the panel has nothing to work with.

Where to find it

View > Machine Learning > Machine Learning. The Selector Panel, where trained selectors are used, is in the same group.

Bag Selection

A checkbox per bag in the current map, labeled with the bag’s name (or “Bag N” if it has none). You can select up to 10 bags at once.

Selecting at least one bag sets the plot axes to the map’s two embedding dimensions; clearing the selection clears them again.

Bag Refinement

Available when exactly two bags are selected. Choose Flat Euclidean from the refinement dropdown and Mantis measures how far each point sits from each bag’s center, then proposes points that would sit better in the other bag.

You get:

• Distance statistics: average, maximum and minimum distance, with a Show Lines / Hide Lines toggle that draws the distances on the map.
• Min Improvement Threshold: the improvement a move has to clear before it is suggested.
• A suggested moves list reading Bag 1 → Bag 2 (0.412), each with its own Move button.
• Refine All: applies every suggested move to the working copy and recomputes the suggestions, without saving.
• Save: applies the moves and writes both bags back to the space.

Selecting any number of bags other than two resets the refinement state.

Flat Euclidean is the only refinement mode implemented.

Feature Selection

Under the bag controls, choose which features the selector trains on:

• Embedding Features: Embedding X and Embedding Y, the map’s two coordinates.
• Additional Features: the numeric metadata columns of the points in your selected bags.
• Feature Statistics: mean, min and max for each feature, computed over the points in the selected bags only, so the numbers move as your selection changes.

Except for the Siamese algorithm, at least one feature must be selected or training fails with “No features selected for training”.

Create Selector

• Selector name: optional. Name it if you want to reuse it; unnamed runs classify without persisting a reusable selector.
• Bag Classifier: the only classifier type on offer.
• Algorithm: SVM, XGBoost, MLP, or Siamese Network (2D Projection).
• Classify Selected Bags: trains the model. The button is disabled until at least one bag is selected.

Choosing Siamese changes the panel: the train button becomes Train Siamese Network, and a Generate 2D Embeddings button appears. Generating writes the network’s 2D output straight into the point coordinates held in the browser, so the map redraws with the learned layout. That is a client-side override of the displayed coordinates, not a new saved map.

Training

Training opens a tracker with a Training Progress bar and a Logs list that walks through data collection, feature preparation, the request, model training and cross-validation, ending with “✓ Training complete!”. The percentages are the panel’s own staged estimate, not epoch-level reporting from the backend.

Start and finish are both written to your User Logs. Dismiss the tracker when it is done.

1. Make at least two bags representing the classes you care about.
2. Open Machine Learning and tick those bags under Bag Selection.
3. Choose features. Embedding X and Y alone often work; add numeric metadata columns to give the model more to go on.
4. Name the selector so you can use it later.
5. Pick an algorithm and click Classify Selected Bags. Watch the tracker.

Using a trained selector

Switch to View > Machine Learning > Selector Panel, headed “Bag Prediction”.

Every selector trained in this space is listed with a checkbox. Show Details expands a report:

• Classifier type, training date, total points, number of bags.
• The selected features and which embeddings were used.
• Mean CV Score and Mean AUC, as percentages.
• The selector’s ROC curve.

Delete removes a selector permanently.

Under Select Bag to Predict, pick either New Bag or one of your existing bags, then:

• Discriminate Bag: runs the selector over the points in the bag you chose and shows a per-point prediction table.
• Create Bags: enabled when New Bag is chosen; runs the same prediction without restricting it to one bag’s points.
• Save Predictions to Bags: appears once results exist, and turns the predictions into real bags, one per predicted class, named after the selector. Points the model assigns to the trailing “none of the above” class are left out.

Limits

• Ten bags maximum per comparison.
• One refinement mode, and it needs exactly two bags.
• Feature statistics and training both use only the points inside the selected bags.
• Training progress is estimated by the panel, so a long run can sit at a high percentage for a while.
• Selectors belong to the space they were trained in.

Read nextBagsEvery workflow on this page starts with bags.