leakage_vergleich.png

def fig_leakage(X_num, y, cv):
    rng = np.random.default_rng(SEED)
    X_aug = np.hstack([StandardScaler().fit_transform(X_num), rng.normal(size=(len(y), 300))])
    leaky = cross_val_score(LogisticRegression(max_iter=1000),
                            SelectKBest(f_classif, k=12).fit_transform(X_aug, y),
                            y, cv=cv, scoring="roc_auc").mean()
    honest = cross_val_score(Pipeline([("s", SelectKBest(f_classif, k=12)),
                                       ("m", LogisticRegression(max_iter=1000))]),
                             X_aug, y, cv=cv, scoring="roc_auc").mean()
    fig, ax = plt.subplots(figsize=(6, 3.4))
    ax.bar(["mit Leakage\n(Auswahl auf allen Daten)", "korrekt\n(Auswahl in der Pipeline)"],
           [leaky, honest], color=[EVENT, PRIMARY], width=0.6)
    for i, v in enumerate([leaky, honest]):
        ax.text(i, v + 0.01, f"{v:.2f}", ha="center", fontweight="bold")
    ax.axhline(0.5, color=SECONDARY, lw=0.8, ls="--")
    ax.set_ylim(0, 1)
    ax.set_ylabel("Kreuzvalidierte ROC-AUC")
    ax.set_title("Data Leakage täuscht eine bessere Modellgüte vor")
    save(fig, ASSETS / "leakage_vergleich.png")