Experimental ML workflow for link prediction

This note documents where the exploratory ML pieces from the original proxfun_full.py prototype now belong.

The package core should remain focused on reusable link prediction metrics:

src/relationalstats/linkprediction/

Experimental workflows should live under:

examples/linkprediction/

Migrated experimental ideas

The original prototype included exploratory utilities and notes for:

• XGBoost-based feature evaluation;
• automatic metric-combination search;
• feature importance for tree models;
• PCA-based feature exploration;
• correlation heatmaps with seaborn;
• AUC distribution plots with matplotlib;
• SHAP-based model explanation.

Those ideas are intentionally not part of the installable core API.

They are represented as optional example code in:

examples/linkprediction/ml_feature_search_xgboost.py

Optional dependencies

Install optional ML and plotting dependencies only when needed:

python -m pip install -e ".[dev,test,plot]"

NOTE: XGBoost is optional. On macOS, XGBoost may require the OpenMP runtime libomp.dylib. The example defaults to a scikit-learn classifier to keep the workflow portable.

Why this is outside the package core

Keeping this workflow outside src/relationalstats/ avoids forcing users to install heavier dependencies such as:

• xgboost
• shap
• seaborn
• matplotlib

It also keeps the public API focused on stable graph metrics rather than notebook-style experimentation.

Recommended execution

python examples/linkprediction/ml_feature_search_xgboost.py

The script uses a small synthetic graph and does not depend on private notebooks, course assignments, or external datasets.