Meaningful Features
Meaningful Features answers one question: how does this group of points differ from everything else in the space? It compares your selection against a sample of the rest of the map, scores every column, flags anomalous points inside the selection, and ranks features by how well they separate the group.
Where to find it
View → Quantitative → Meaningful Features. You can also right-click the map with points selected and pick Feature Analysis under Analyze; that entry is disabled until at least five points are selected and the space has numeric columns. The panel’s own header reads “Feature Analysis”, though the View menu lists it as Meaningful Features.
What it reads
The current selection, not a bag. This is the exception in the Quantitative group: lasso some points, click a cluster, or select a bag’s points and the panel picks them up directly. With fewer than five points selected, it shows “Select at least 5 points” and the Analyze Selection button stays disabled.
Opening the panel with five or more points already selected runs the analysis once automatically. After that, press Analyze Selection to rerun against the current selection, or Clear to empty all three tabs. The header shows how many points the last run covered.
The three tabs
Group Characteristics
Every column scored for how strongly it separates your selection from the rest of the space.
Numeric columns get a table of Feature, Effect Size (d), Sel. Mean, Rest Mean, and p-value. The effect size is Cohen’s d, green when the selection sits above the rest and red when it sits below, with a bar scaled so that |d| = 2 fills it. The p-value comes from Welch’s t-test and its badge is green under 0.05, amber under 0.1, grey above.
Categorical columns get Cramer’s V, the selection’s mode, the rest’s mode, and a chi-squared p-value.
The banner at the top counts the features that pass the built-in thresholds, |d| > 0.5 for numeric and V > 0.3 for categorical. When nothing passes, it says so and still lists every feature underneath.
Outliers
Points inside the selection that do not look like the others. The score is an ensemble of Isolation Forest, Local Outlier Factor, and per-column z-scores, averaged and normalized; the top decile is flagged. Each card shows the point’s id prefix, its combined score, the columns where its z-score exceeds 2.5, and bars for its five largest per-column deviations. Click a card to make that point the current selection so the Inspector and Reader follow it.
Outlier detection needs at least five points with numeric data; below that it returns nothing.
Feature Relevance
The top 15 columns ranked by a combined score of mutual information, ANOVA F-statistic, and random-forest importance, each normalized before averaging. A badge marks each row num or cat. Underneath, Chart Suggestions proposes chart types and field pairs worth plotting, each with a one-line reason.
How to use it
- Select at least five points on the map, or select a bag.
- Open the panel. It analyzes the selection on open.
- Read Group Characteristics for the columns that separate the group, sorted by effect size.
- Check Outliers to find the points dragging your group’s averages around, and click one to inspect it.
- Take a suggestion from Feature Relevance into Plot Controls to see the separation.
Limits
- Group Characteristics and Feature Relevance need a comparison set. If the rest of the map cannot be sampled, those two tabs stay empty while Outliers still works.
- Change the selection and the results do not follow. Press Analyze Selection again.
- Chart Suggestions are text. Clicking one does nothing; set the axes yourself in Plot Controls.
- Nothing here is exported or saved with the space.
- There is no agent button in this panel.