@nahisaho/satori 0.1.0

package/src/.github/skills/scientific-ml-classification/SKILL.md

@@ -0,0 +1,265 @@
+---
+name: scientific-ml-classification
+description: |
+  機械学習分類パイプラインのスキル。複数の分類モデル（Logistic Regression, Random Forest,
+  SVM, XGBoost）を StratifiedKFold 交差検証で比較し、ROC 曲線・混同行列で評価する際に使用。
+  Scientific Skills Exp-03, 05 で確立したパターン。
+---
+
+# Scientific ML Classification Pipeline
+
+バイナリ/マルチクラス分類タスクのための統一 ML パイプライン。
+StratifiedKFold 交差検証で複数モデルを公平に比較し、ROC-AUC と混同行列で評価する。
+
+## When to Use
+
+- カテゴリ予測タスク（二値分類・多クラス分類）
+- がん分類、毒性予測、材料分類などの判別問題
+- 複数分類モデルの性能比較が必要なとき
+
+## Quick Start
+
+## 標準パイプライン
+
+### 1. モデル定義
+
+```python
+from sklearn.linear_model import LogisticRegression
+from sklearn.ensemble import RandomForestClassifier, GradientBoostingClassifier
+from sklearn.svm import SVC
+from sklearn.model_selection import StratifiedKFold, cross_val_score
+from sklearn.preprocessing import StandardScaler
+from sklearn.metrics import (accuracy_score, precision_score, recall_score,
+                             f1_score, roc_auc_score, roc_curve,
+                             confusion_matrix, classification_report)
+
+MODEL_DEFS = {
+    "Logistic Regression": LogisticRegression(max_iter=1000, random_state=42),
+    "Random Forest": RandomForestClassifier(n_estimators=200, max_depth=15,
+                                            random_state=42),
+    "Gradient Boosting": GradientBoostingClassifier(n_estimators=200,
+                                                     random_state=42),
+    "SVM": SVC(kernel="rbf", probability=True, random_state=42),
+}
+```
+
+### 2. StratifiedKFold 学習 & 評価
+
+```python
+import numpy as np
+import pandas as pd
+
+def train_evaluate_classifiers(X_train, X_test, y_train, y_test,
+                                model_defs=None, n_splits=5):
+    """全分類モデルを StratifiedKFold で評価する。"""
+    if model_defs is None:
+        model_defs = MODEL_DEFS
+
+    scaler = StandardScaler()
+    X_train_sc = scaler.fit_transform(X_train)
+    X_test_sc = scaler.transform(X_test)
+
+    skf = StratifiedKFold(n_splits=n_splits, shuffle=True, random_state=42)
+    results = []
+    trained_models = {}
+
+    for name, model in model_defs.items():
+        import copy
+        m = copy.deepcopy(model)
+        m.fit(X_train_sc, y_train)
+        y_pred = m.predict(X_test_sc)
+        y_proba = m.predict_proba(X_test_sc)[:, 1] if hasattr(m, "predict_proba") else None
+
+        cv_scores = cross_val_score(copy.deepcopy(model), X_train_sc, y_train,
+                                    cv=skf, scoring="accuracy")
+
+        results.append({
+            "Model": name,
+            "Accuracy": accuracy_score(y_test, y_pred),
+            "Precision": precision_score(y_test, y_pred, average="weighted"),
+            "Recall": recall_score(y_test, y_pred, average="weighted"),
+            "F1": f1_score(y_test, y_pred, average="weighted"),
+            "ROC_AUC": roc_auc_score(y_test, y_proba) if y_proba is not None else np.nan,
+            "CV_Accuracy_mean": cv_scores.mean(),
+            "CV_Accuracy_std": cv_scores.std(),
+        })
+        trained_models[name] = {"model": m, "y_pred": y_pred, "y_proba": y_proba}
+
+    results_df = pd.DataFrame(results)
+    results_df.to_csv("results/classification_metrics.csv", index=False)
+    return results_df, trained_models
+```
+
+### 3. ROC 曲線
+
+```python
+import matplotlib.pyplot as plt
+
+def plot_roc_curves(y_test, trained_models, figsize=(8, 8)):
+    """全モデルの ROC 曲線を重ねて描画する。"""
+    fig, ax = plt.subplots(figsize=figsize)
+
+    for name, info in trained_models.items():
+        if info["y_proba"] is not None:
+            fpr, tpr, _ = roc_curve(y_test, info["y_proba"])
+            auc = roc_auc_score(y_test, info["y_proba"])
+            ax.plot(fpr, tpr, linewidth=2, label=f"{name} (AUC={auc:.3f})")
+
+    ax.plot([0, 1], [0, 1], "k--", linewidth=1)
+    ax.set_xlabel("False Positive Rate")
+    ax.set_ylabel("True Positive Rate")
+    ax.set_title("ROC Curves", fontweight="bold")
+    ax.legend()
+    plt.tight_layout()
+    plt.savefig("figures/roc_curves.png", dpi=300, bbox_inches="tight")
+    plt.close()
+```
+
+### 4. 混同行列
+
+```python
+import seaborn as sns
+
+def plot_confusion_matrices(y_test, trained_models, class_names=None,
+                            ncols=2, figsize=(14, 12)):
+    """全モデルの混同行列をグリッド表示する。"""
+    n_models = len(trained_models)
+    nrows = (n_models + ncols - 1) // ncols
+    fig, axes = plt.subplots(nrows, ncols, figsize=figsize)
+    axes = axes.flatten()
+
+    for i, (name, info) in enumerate(trained_models.items()):
+        cm = confusion_matrix(y_test, info["y_pred"])
+        sns.heatmap(cm, annot=True, fmt="d", cmap="Blues", ax=axes[i],
+                    xticklabels=class_names, yticklabels=class_names)
+        axes[i].set_title(name, fontweight="bold")
+        axes[i].set_xlabel("Predicted")
+        axes[i].set_ylabel("Actual")
+
+    for j in range(i + 1, len(axes)):
+        axes[j].set_visible(False)
+
+    plt.tight_layout()
+    plt.savefig("figures/confusion_matrices.png", dpi=300, bbox_inches="tight")
+    plt.close()
+```
+
+### 5. Precision-Recall 曲線
+
+```python
+from sklearn.metrics import precision_recall_curve, average_precision_score
+
+def plot_precision_recall_curves(y_test, trained_models, figsize=(8, 8)):
+    """全モデルの Precision-Recall 曲線を重ねて描画する。
+    クラス不均衡データでは ROC よりも PR 曲線が適切な評価指標となる。"""
+    fig, ax = plt.subplots(figsize=figsize)
+
+    for name, info in trained_models.items():
+        if info["y_proba"] is not None:
+            precision, recall, _ = precision_recall_curve(y_test, info["y_proba"])
+            ap = average_precision_score(y_test, info["y_proba"])
+            ax.plot(recall, precision, linewidth=2,
+                    label=f"{name} (AP={ap:.3f})")
+
+    # ベースライン（陽性率）
+    baseline = y_test.mean()
+    ax.axhline(baseline, color="gray", linestyle="--", linewidth=1,
+               label=f"Baseline ({baseline:.3f})")
+
+    ax.set_xlabel("Recall")
+    ax.set_ylabel("Precision")
+    ax.set_title("Precision-Recall Curves", fontweight="bold")
+    ax.legend()
+    ax.set_xlim([0, 1])
+    ax.set_ylim([0, 1.05])
+    plt.tight_layout()
+    plt.savefig("figures/precision_recall_curves.png", dpi=300,
+                bbox_inches="tight")
+    plt.close()
+```
+
+### 6. Partial Dependence Plot (PDP)
+
+```python
+from sklearn.inspection import PartialDependenceDisplay
+
+def plot_partial_dependence(model, X_train_scaled, feature_names,
+                             top_n=6, figsize=(14, 8)):
+    """重要な特徴量の部分依存プロットを描画する。
+    モデルの予測が各特徴量に対してどのように変化するかを可視化。"""
+    fig, ax = plt.subplots(figsize=figsize)
+    features_idx = list(range(min(top_n, len(feature_names))))
+    PartialDependenceDisplay.from_estimator(
+        model, X_train_scaled, features=features_idx,
+        feature_names=feature_names,
+        ax=ax, grid_resolution=50
+    )
+    plt.suptitle("Partial Dependence Plots", fontweight="bold", y=1.02)
+    plt.tight_layout()
+    plt.savefig("figures/partial_dependence.png", dpi=300, bbox_inches="tight")
+    plt.close()
+```
+
+### 7. Volcano Plot（差次発現/差次特徴量）
+
+```python
+from scipy import stats
+
+def volcano_plot(df, group_col, value_cols, group1, group2,
+                 fc_threshold=1.0, p_threshold=0.05, figsize=(10, 8)):
+    """Fold Change と p 値による Volcano Plot を描画する。"""
+    results = []
+    g1 = df[df[group_col] == group1]
+    g2 = df[df[group_col] == group2]
+
+    for col in value_cols:
+        stat, pval = stats.mannwhitneyu(g1[col].dropna(), g2[col].dropna(),
+                                        alternative="two-sided")
+        fc = g2[col].mean() - g1[col].mean()  # or log2FC
+        results.append({"Feature": col, "log2FC": fc, "pvalue": pval,
+                        "neg_log10p": -np.log10(pval + 1e-300)})
+
+    vdf = pd.DataFrame(results)
+
+    fig, ax = plt.subplots(figsize=figsize)
+    # 色分け: significant up / down / ns
+    sig_up = (vdf["log2FC"] > fc_threshold) & (vdf["pvalue"] < p_threshold)
+    sig_down = (vdf["log2FC"] < -fc_threshold) & (vdf["pvalue"] < p_threshold)
+    ns = ~(sig_up | sig_down)
+
+    ax.scatter(vdf.loc[ns, "log2FC"], vdf.loc[ns, "neg_log10p"],
+              c="gray", alpha=0.5, s=20, label="NS")
+    ax.scatter(vdf.loc[sig_up, "log2FC"], vdf.loc[sig_up, "neg_log10p"],
+              c="red", alpha=0.7, s=30, label="Up")
+    ax.scatter(vdf.loc[sig_down, "log2FC"], vdf.loc[sig_down, "neg_log10p"],
+              c="blue", alpha=0.7, s=30, label="Down")
+    ax.axhline(-np.log10(p_threshold), color="gray", linestyle="--")
+    ax.axvline(fc_threshold, color="gray", linestyle="--")
+    ax.axvline(-fc_threshold, color="gray", linestyle="--")
+    ax.set_xlabel("log₂ Fold Change")
+    ax.set_ylabel("-log₁₀(p-value)")
+    ax.set_title("Volcano Plot", fontweight="bold")
+    ax.legend()
+    plt.tight_layout()
+    plt.savefig("figures/volcano_plot.png", dpi=300, bbox_inches="tight")
+    plt.close()
+    return vdf
+```
+
+## References
+
+### Output Files
+
+| ファイル | 形式 |
+|---|---|
+| `results/classification_metrics.csv` | CSV |
+| `figures/roc_curves.png` | PNG |
+| `figures/precision_recall_curves.png` | PNG |
+| `figures/confusion_matrices.png` | PNG |
+| `figures/partial_dependence.png` | PNG |
+| `figures/volcano_plot.png` | PNG |
+
+#### 参照実験
+
+- **Exp-03**: がん遺伝子発現の 4 モデル分類 + ROC + 混同行列
+- **Exp-05**: 毒性予測の二値分類 + Volcano Plot + PR 曲線 + PDP

package/src/.github/skills/scientific-ml-regression/SKILL.md

@@ -0,0 +1,215 @@
+---
+name: scientific-ml-regression
+description: |
+  マルチターゲット回帰モデルの学習・評価・比較スキル。複数の回帰モデル（Ridge, Lasso,
+  Random Forest, Gradient Boosting, Extra Trees）を KFold 交差検証で比較する際に使用。
+  Scientific Skills Exp-12, 13 で確立したパターン。
+---
+
+# Scientific ML Regression Pipeline
+
+複数の回帰モデルを統一フレームワークで学習・評価・比較するパイプライン。
+マルチターゲット回帰（複数の出力変数を同時に予測）に対応。
+
+## When to Use
+
+- 連続値の予測タスク
+- 複数モデルの性能比較が必要なとき
+- マルチターゲット回帰（複数出力変数の同時予測）
+- 交差検証による汎化性能の評価
+
+## Quick Start
+
+## 標準パイプライン
+
+### 1. モデル定義
+
+```python
+from sklearn.linear_model import Ridge, Lasso, ElasticNet
+from sklearn.ensemble import (RandomForestRegressor,
+                              GradientBoostingRegressor,
+                              ExtraTreesRegressor)
+from sklearn.model_selection import KFold, cross_val_score
+from sklearn.preprocessing import StandardScaler
+from sklearn.metrics import r2_score, mean_squared_error, mean_absolute_error
+import numpy as np
+import pandas as pd
+
+MODEL_DEFS = {
+    "Ridge": Ridge(alpha=1.0),
+    "Lasso": Lasso(alpha=0.1, max_iter=10000),
+    "Random Forest": RandomForestRegressor(n_estimators=200, max_depth=15,
+                                           random_state=42),
+    "Gradient Boosting": GradientBoostingRegressor(n_estimators=200,
+                                                    random_state=42),
+    "Extra Trees": ExtraTreesRegressor(n_estimators=200, max_depth=15,
+                                       random_state=42),
+}
+```
+
+### 2. マルチターゲット学習 & 評価
+
+```python
+def train_evaluate_models(X_train, X_test, y_train, y_test,
+                          target_names, model_defs=None, n_splits=5):
+    """
+    全モデル × 全ターゲットの学習・評価を一括実行する。
+    Returns: results_df (モデル×ターゲットの評価指標), best_models dict
+    """
+    if model_defs is None:
+        model_defs = MODEL_DEFS
+
+    scaler = StandardScaler()
+    X_train_sc = scaler.fit_transform(X_train)
+    X_test_sc = scaler.transform(X_test)
+
+    results = []
+    best_models = {}
+    kf = KFold(n_splits=n_splits, shuffle=True, random_state=42)
+
+    for target in target_names:
+        y_tr = y_train[target].values
+        y_te = y_test[target].values
+        best_r2 = -np.inf
+
+        for name, model in model_defs.items():
+            import copy
+            m = copy.deepcopy(model)
+            m.fit(X_train_sc, y_tr)
+            y_pred = m.predict(X_test_sc)
+
+            r2 = r2_score(y_te, y_pred)
+            rmse = np.sqrt(mean_squared_error(y_te, y_pred))
+            mae = mean_absolute_error(y_te, y_pred)
+
+            # 交差検証
+            cv_scores = cross_val_score(
+                copy.deepcopy(model), X_train_sc, y_tr,
+                cv=kf, scoring="neg_root_mean_squared_error"
+            )
+            cv_rmse_mean = -cv_scores.mean()
+            cv_rmse_std = cv_scores.std()
+
+            results.append({
+                "Target": target,
+                "Model": name,
+                "R2": r2,
+                "RMSE": rmse,
+                "MAE": mae,
+                "CV_RMSE_mean": cv_rmse_mean,
+                "CV_RMSE_std": cv_rmse_std,
+            })
+
+            if r2 > best_r2:
+                best_r2 = r2
+                best_models[target] = {"model": m, "name": name, "r2": r2}
+
+    results_df = pd.DataFrame(results)
+    results_df.to_csv("results/model_metrics.csv", index=False)
+    return results_df, best_models
+```
+
+### 3. モデル比較可視化（R² バーチャート）
+
+```python
+import matplotlib.pyplot as plt
+import seaborn as sns
+
+def plot_model_comparison_r2(results_df, figsize=(14, 6)):
+    """R² の grouped bar chart を描画する。"""
+    pivot = results_df.pivot(index="Target", columns="Model", values="R2")
+    ax = pivot.plot(kind="bar", figsize=figsize, colormap="Set2", edgecolor="black")
+    ax.set_ylabel("R² Score")
+    ax.set_title("Model Comparison: R² by Target", fontweight="bold")
+    ax.set_ylim(0, 1.05)
+    ax.legend(title="Model", bbox_to_anchor=(1.05, 1))
+    ax.axhline(y=0.8, color="gray", linestyle="--", alpha=0.5, label="R²=0.8")
+    plt.xticks(rotation=45, ha="right")
+    plt.tight_layout()
+    plt.savefig("figures/model_comparison_r2.png", dpi=300, bbox_inches="tight")
+    plt.close()
+```
+
+### 4. Actual vs Predicted プロット
+
+```python
+def plot_actual_vs_predicted(X_test_sc, y_test, best_models, target_names,
+                             ncols=3, figsize=(18, 12)):
+    """最良モデルの実測値 vs 予測値プロットを描画する。"""
+    nrows = (len(target_names) + ncols - 1) // ncols
+    fig, axes = plt.subplots(nrows, ncols, figsize=figsize)
+    axes = axes.flatten()
+
+    for i, target in enumerate(target_names):
+        m = best_models[target]["model"]
+        y_pred = m.predict(X_test_sc)
+        y_true = y_test[target].values
+        r2 = best_models[target]["r2"]
+
+        axes[i].scatter(y_true, y_pred, alpha=0.5, s=20, edgecolors="k", linewidth=0.3)
+        lims = [min(y_true.min(), y_pred.min()), max(y_true.max(), y_pred.max())]
+        axes[i].plot(lims, lims, "r--", linewidth=1.5)
+        axes[i].set_xlabel("Actual")
+        axes[i].set_ylabel("Predicted")
+        axes[i].set_title(f"{target}\n({best_models[target]['name']}, R²={r2:.3f})",
+                         fontsize=10, fontweight="bold")
+
+    for j in range(i + 1, len(axes)):
+        axes[j].set_visible(False)
+
+    plt.tight_layout()
+    plt.savefig("figures/actual_vs_predicted.png", dpi=300, bbox_inches="tight")
+    plt.close()
+```
+
+### 5. レーダーチャート比較（Exp-05 パターン）
+
+```python
+def plot_radar_comparison(results_df, target_name, figsize=(8, 8)):
+    """指定ターゲットのモデル比較レーダーチャートを描画する。"""
+    subset = results_df[results_df["Target"] == target_name]
+    metrics = ["R2", "1-RMSE_norm", "1-MAE_norm", "Stability"]
+
+    # 正規化（全指標を 0-1, 高い方が良い に統一）
+    subset = subset.copy()
+    for col in ["RMSE", "MAE"]:
+        max_val = subset[col].max()
+        subset[f"1-{col}_norm"] = 1 - subset[col] / max_val if max_val > 0 else 1
+    subset["Stability"] = 1 - subset["CV_RMSE_std"] / subset["CV_RMSE_mean"]
+
+    N = len(metrics)
+    angles = np.linspace(0, 2 * np.pi, N, endpoint=False).tolist()
+    angles += angles[:1]
+
+    fig, ax = plt.subplots(figsize=figsize, subplot_kw=dict(polar=True))
+
+    for _, row in subset.iterrows():
+        values = [row[m] for m in metrics] + [row[metrics[0]]]
+        ax.plot(angles, values, "o-", linewidth=2, label=row["Model"])
+        ax.fill(angles, values, alpha=0.1)
+
+    ax.set_xticks(angles[:-1])
+    ax.set_xticklabels(metrics)
+    ax.set_title(f"Radar: {target_name}", fontweight="bold")
+    ax.legend(loc="upper right", bbox_to_anchor=(1.3, 1.1))
+    plt.tight_layout()
+    plt.savefig(f"figures/radar_{target_name}.png", dpi=300, bbox_inches="tight")
+    plt.close()
+```
+
+## References
+
+### Output Files
+
+| ファイル | 形式 |
+|---|---|
+| `results/model_metrics.csv` | CSV |
+| `figures/model_comparison_r2.png` | PNG |
+| `figures/actual_vs_predicted.png` | PNG |
+| `figures/radar_*.png` | PNG |
+
+#### 参照実験
+
+- **Exp-12**: 6 モデル × 4 ターゲット（エッチング品質特性）
+- **Exp-13**: 5 モデル × 6 ターゲット（PSP 構造・物性）
+- **Exp-05**: レーダーチャートによるモデル比較