auc_vergleich.png

def fig_auc_vergleich(X, y) -> None:
    """Bar chart comparing 5-fold CV AUC: LR vs GB vs MLP."""
    cv = StratifiedKFold(n_splits=5, shuffle=True, random_state=SEED)

    def lr_pipe():
        return Pipeline([("pre", build_preprocessor()),
                         ("clf", LogisticRegression(max_iter=1000,
                                                    class_weight="balanced",
                                                    random_state=SEED))])

    def gb_pipe():
        return Pipeline([("pre", build_preprocessor()),
                         ("clf", GradientBoostingClassifier(n_estimators=200,
                                                             max_depth=3,
                                                             random_state=SEED))])

    def mlp_pipe():
        # early_stopping=False for CV: avoids shrinking each fold's training set
        # max_iter=500 prevents ConvergenceWarnings on this small dataset
        return Pipeline([("pre", build_preprocessor()),
                         ("mlp", MLPClassifier(hidden_layer_sizes=(64, 32),
                                               alpha=0.01,
                                               early_stopping=False,
                                               max_iter=500,
                                               random_state=SEED))])

    models = [
        ("Logistische\nRegression", lr_pipe(), PRIMARY),
        ("Gradient\nBoosting", gb_pipe(), PALETTE[2]),
        ("MLP\n(Neuronales Netz)", mlp_pipe(), EVENT),
    ]

    names, aucs, colors = [], [], []
    for name, pipe, color in models:
        scores = cross_val_score(pipe, X, y, cv=cv, scoring="roc_auc")
        names.append(name)
        aucs.append(scores.mean())
        colors.append(color)
        print(f"  {name.replace(chr(10), ' ')}: AUC={scores.mean():.3f}")

    fig, ax = plt.subplots(figsize=(6.5, 4))
    bars = ax.bar(names, aucs, color=colors, width=0.55)
    for bar, auc in zip(bars, aucs):
        ax.text(bar.get_x() + bar.get_width() / 2, auc + 0.005,
                f"{auc:.3f}", ha="center", fontweight="bold", fontsize=11)

    ax.axhline(0.5, color=SECONDARY, lw=0.8, ls="--")
    ax.set_ylim(0, 1)
    ax.set_ylabel("Kreuzvalidierte ROC-AUC (5-fach, stratifiziert)")

    # Data-driven verdict — NEVER hardcode who "wins": derive it from the
    # computed aucs so the title can't drift out of sync with the bars.
    order = sorted(zip(names, aucs), key=lambda t: -t[1])
    best_name, best_auc = order[0]
    runner_up_name, runner_up_auc = order[1]
    gap = best_auc - runner_up_auc
    best_flat = best_name.replace("\n", " ")
    runner_up_flat = runner_up_name.replace("\n", " ")
    verdict = (f"{best_flat} liegt vorn" if gap > 0.01
               else f"{best_flat} und {runner_up_flat} liegen gleichauf")

    ax.set_title("AUC-Vergleich: Logistische Regression vs. Gradient Boosting vs. MLP\n"
                 f"(synthetische Kohorte, n≈{len(X)} — {verdict})")

    save(fig, ASSETS / "auc_vergleich.png")