rsf-grid-search

A small grid search over RandomSurvivalForest hyperparameters (min_samples_leaf, max_features, max_depth), scored by Harrell's concordance index on a held-out validation split.

Random survival forests have no closed-form regularization path the way a penalized Cox model does, so their main regularization knobs are tuned by grid search instead. events_per_leaf is reported alongside the C-index for each combination as a rough diagnostic: a leaf with too few events gives an unstable local Kaplan-Meier estimate, the survival-forest analogue of too few events per parameter in a Cox model.

Install

pip install rsf-grid-search

Usage

from rsf_grid_search import grid_rsf

results = grid_rsf(X, y, n_sub=30000, n_trees=300, seed=1)
print(results.head())

X is a pandas.DataFrame design matrix and y is a survival target in scikit-survival's structured-array format (fields "event", "time", e.g. built with sksurv.util.Surv.from_arrays).

If the training split is larger than n_sub, it is subsampled (stratified on the event indicator) before the grid search, to keep each fit's runtime bounded. The result is a DataFrame, one row per grid point, sorted by descending validation C-index, with columns leaf, max_features, depth, events_per_leaf, c_index, secs.

The default grid is min_samples_leaf in (30, 60, 100, 150, 250), max_features in ("sqrt", 0.2, 0.35, 0.6), and max_depth in (None, 10). Override any of them:

results = grid_rsf(
    X, y,
    leaf_grid=(20, 50, 100),
    max_features_grid=("sqrt", 0.5),
    depth_grid=(None,),
)

License

MIT