Function mdi_importance

python/openquant/feature_diagnostics.py:260–304

(
    X: Sequence[Sequence[float]],
    y: Sequence[float],
    feature_names: Sequence[str] | None = None,
    sample_weight: Sequence[float] | None = None,
    n_estimators: int = 32,
    seed: int = 42,
)

def _importance_table(feature_names: Sequence[str], per_feature_values: Sequence[Sequence[float]]) -> pl.DataFrame:
    rows = []
    for name, vals in zip(feature_names, per_feature_values):
        rows.append(
            {
                "feature": name,
                "mean": _mean(vals),
                "std": _std(vals),
                "stderr": _std(vals) / sqrt(len(vals)) if vals else 0.0,
            }
        )
    return pl.DataFrame(rows).sort("mean", descending=True)


def mdi_importance(
    X: Sequence[Sequence[float]],
    y: Sequence[float],
    feature_names: Sequence[str] | None = None,
    sample_weight: Sequence[float] | None = None,
    n_estimators: int = 32,
    seed: int = 42,
) -> dict[str, object]:
    x = _as_matrix(X)
    yv = _as_vector(y, len(x))
    names = _feature_names(len(x[0]), feature_names)
    weights = _sample_weight(sample_weight, len(x))

    if n_estimators < 2:
        raise ValueError("n_estimators must be >= 2")

    import random

    rng = random.Random(seed)
    per_feature: list[list[float]] = [[] for _ in names]

    for _ in range(n_estimators):
        idx = [rng.randrange(len(x)) for _ in range(len(x))]
        xb = [x[i] for i in idx]
        yb = [yv[i] for i in idx]
        wb = [weights[i] for i in idx] if weights is not None else None

        model = _fit_linear_probability_model(xb, yb, wb)
        raw = [abs(v) for v in model.coeffs]
        denom = sum(raw)
        norm = [v / denom if denom > 0 else 0.0 for v in raw]
        for j, v in enumerate(norm):
            per_feature[j].append(v)