Orange3-Robust 0.1.0

Robust linear regression (Huber, LAD, RANSAC, Theil–Sen) for Orange.

Robust Regression for Orange

This Orange add-on adds two widgets for learning how regression behaves when a data set contains outliers.

Ordinary linear regression fits the line or plane that minimizes squared error. That makes large errors very influential: a few unusual rows can pull the model away from the main pattern. Robust regression methods reduce that effect in different ways.

Widgets

Robust Regression

Fits a robust linear regression learner. The widget currently includes:

• Huber: behaves like ordinary least squares for small residuals, but gives less influence to large residuals.
• Least Absolute Deviation: fits by absolute error when the quantile is 0.5, so it targets the conditional median rather than the conditional mean. You can change the quantile to fit lower or upper conditional quantiles.
• RANSAC: repeatedly fits candidate models from small subsets, then keeps the model supported by the largest group of inliers.
• Theil-Sen: estimates slopes from many small subsets and combines them in a way that is resistant to outliers.

The widget outputs a learner for Test and Score and Predictions. If you connect a data set with a continuous target variable, it also outputs a fitted model, a coefficient table, and an annotated data table.

Robust Scale

Centers continuous features by the median and scales them by a quantile range, usually the interquartile range. This is useful when large values should not determine the scale of a feature.

Suggested Workflow

1. Add File or Datasets.
2. Add Select Columns and choose a continuous target variable.
3. Add Robust Regression.
4. Connect Robust Regression to Test and Score or Predictions.
5. Connect the Annotated Data output to Data Table to inspect residuals.
6. Connect the Coefficients output to Data Table to inspect fitted terms.

The demo workflow in docs/examples/robust-regression-demo.ows shows this layout:

One set of widget settings looks like this:

What To Inspect

The Annotated Data output adds:

• the model prediction for each row;
• the residual, computed as actual value minus prediction;
• a RANSAC inlier flag when RANSAC is selected.

The Coefficients output lists the intercept and fitted coefficients when the selected method exposes them. These are useful for comparing robust fits with ordinary linear regression on the same data.

Notes

Do not use Robust Scale before Robust Regression if the regression widget's internal scaling checkbox is also enabled, because that scales continuous features twice.

Developer setup and release notes are in DEVELOPMENT.md.