fix(eval): fix embedding averaging in privacy metrics

[nina-xu]

Summary

Per this GitLab issue, we are averaging column embeddings incorrectly in the two privacy metrics, AIP and MIP. Later columns get a higher weight in the averaging. Taking Aaron's suggestion in the ticket for an equal weighting.

In addition, I DRY'ed up the code, taking common methods out into privacy_metric_utils.py, because both metrics utilize the same way of averaging text column embeddings.

I also added some tests to illustrate that the new way of averaging gives each column equal weight.

Pre-Review Checklist

Ensure that the following pass:

• make format && make lint or via prek validation.
• make test passes locally
• make test-e2e passes locally
• make test-ci-container passes locally (recommended)

Pre-Merge Checklist

• New or updated tests for any fix or new behavior
• Updated documentation for new features and behaviors, including docstrings for API docs.

Testing Plan

• Add unit tests
• Take one dataset, shuffle around columns and calculate privacy scores to show the effect before & after
• Run slurm experiments to compare the scores before & after, so that we are aware of the new baseline
• We hardly have datasets with 3 or more text columns, so it's hard to demonstrate the change in action. Ran the "short" dataset group; plus dolly (3 text columns) to demonstrate that things are not breaking (wandb). More specifically, dolly has 10 MIA and 10 for AIA for the 3 text columns, so we wouldn't be able to observe the difference from this dataset; patient_events also has 3 text columns, but it also has an AIA of 10

Note

Closes #203

[Copilot]

Pull request overview

Fixes incorrect column-embedding averaging in the AIP (attribute inference) and MIP (membership inference) privacy metrics by switching from sequential pairwise averaging (which overweighted later columns) to an equal-weight mean across all text columns, and factors shared logic into a new utility module.

Changes:

• Introduce privacy_metric_utils.py with shared helpers for identifying text fields, splitting tabular/text columns, and embedding+averaging text columns via np.stack/np.mean.
• Update AIP/MIP components to use the shared helpers and to reuse a single SentenceTransformer instance per metric invocation.
• Add unit tests covering the corrected equal-weight embedding averaging behavior.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.

File	Description
tests/evaluation/components/test_privacy_metric_utils.py	Adds regression/unit tests validating equal-weight embedding averaging and dataframe splitting.
src/nemo_safe_synthesizer/evaluation/components/privacy_metric_utils.py	New shared utilities for text-field detection and embedding aggregation.
src/nemo_safe_synthesizer/evaluation/components/membership_inference_protection.py	Replaces local helpers with shared utilities; fixes embedding averaging logic in MIP path.
src/nemo_safe_synthesizer/evaluation/components/attribute_inference_protection.py	Replaces local helpers with shared utilities; fixes embedding averaging logic in AIP path.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 5 out of 5 changed files in this pull request and generated no new comments.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.