[SPARK-14604] Modify design of ML model summaries

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Incomplete
• Component/s: ML
• Labels: bulk-closed
• Assignee: Unassigned
• Reporter: Joseph K. Bradley
• Created: 13/Apr/16 18:13
• Updated: 08/Oct/19 05:41
• Resolved: 08/Oct/19 05:41

Description

Several spark.ml models now have summaries containing evaluation metrics and training info:

• LinearRegressionModel
• LogisticRegressionModel
• GeneralizedLinearRegressionModel

These summaries have unfortunately been added in an inconsistent way. I propose to reorganize them to have:

• For each model, 1 summary (without training info) and 1 training summary (with info from training). The non-training summary can be produced for a new dataset via evaluate.
• A summary should not store the model itself as a public field.
• A summary should provide a transient reference to the dataset used to produce the summary.

This task will involve reorganizing the GLM summary (which lacks a training/non-training distinction) and deprecating the model method in the LinearRegressionSummary.