[ENH] Feature as Predictor

[janezd]

Issue

Closes #6813.

Description of changes

A widget that "predicts" classes from a single column. I have several questions.

• The widget offers
  • discrete columns whose values are the same, or a subset of, class values,
  • and numeric columns if the class is binary.
• When using numeric columns, its values are used as probabilities of class with index 1 (in which case they must be between 0 and 1) or mapped through logistic function with user-specified offset and coefficient.
• Widget also outputs a model, so it can be fed into Test Learner and compared with other models.

I've put the widget into category Evaluate. This is not a model but rather a trick to turn a Table into Evaluation Results, hence it belongs there because any user interested in this transformation would look for a widget in this category.

Ideas from the discussion:

• Replace radios with a check box to apply log reg. The check box is disabled and checked when the column contains values outside 0 - 1.
• Remove line edits and always compute logistic regression
• Support numeric outcomes; offer fitting with linear regression

Includes

• Code changes
• Tests
• Documentation

[codecov]

Codecov Report

Attention: Patch coverage is 98.90110% with 2 lines in your changes missing coverage. Please review.

Project coverage is 88.29%. Comparing base (70ebf27) to head (62d47f2).
Report is 153 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #6852      +/-   ##
==========================================
+ Coverage   88.27%   88.29%   +0.02%     
==========================================
  Files         326      328       +2     
  Lines       71137    71319     +182     
==========================================
+ Hits        62793    62972     +179     
- Misses       8344     8347       +3     

[Copilot]

Pull Request Overview

This PR introduces a new widget, Feature as Predictor, which repurposes a table column (numeric or discrete) to generate evaluation results and outputs a corresponding model. Key changes include new i18n message entries, widget UI and behavior modifications in owfeatureaspredictor.py along with extensive test coverage, and updates to modelling and classification modules to support the new widget.

Reviewed Changes

Copilot reviewed 9 out of 9 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
i18n/si/msgs.jaml	Added new translation messages for column learner/model errors.
Orange/widgets/evaluate/owfeatureaspredictor.py	Implemented the new widget with control updates and commit logic.
Orange/widgets/evaluate/tests/test_owfeatureaspredictor.py	Added tests to verify behavior and UI interaction of the widget.
Orange/modelling/column.py	Introduced ColumnLearner/ColumnModel with logistic and linear paths.
Orange/tests/test_classification.py	Updated tests to account for ColumnLearner behavior changes.
Orange/modelling/tests/test_column.py	Added tests validating column modelling functionality.
Orange/classification/tests/test_column.py	Added tests ensuring ColumnClassifier handles mapping and predictions.
Orange/modelling/init.py and classification/init.py	Updated exports to include new column modules.

Comments suppressed due to low confidence (1)

Orange/modelling/column.py:99

• The criteria for setting the 'value_mapping' in ColumnModel relies on an implicit slice comparison between 'class_var.values' and 'column.values', which may be fragile if the ordering or lengths differ. Consider adding a clarifying comment or refactoring this logic to explicitly document the intended mapping behavior.

if (column.is_discrete and class_var.values[:len(column.values)] != column.values):