初始化

app.py

@@ -0,0 +1,202 @@
+import gradio as gr
+import pandas as pd
+import numpy as np
+import matplotlib.pyplot as plt
+import seaborn as sns
+from sklearn.model_selection import train_test_split
+from sklearn.preprocessing import StandardScaler
+from sklearn.tree import DecisionTreeClassifier
+from sklearn.svm import SVC
+import xgboost as xgb
+import statsmodels.api as sm
+from sklearn.metrics import confusion_matrix, classification_report, roc_auc_score, accuracy_score
+import warnings
+import json
+
+# --- 初始設定與資料載入 ---
+warnings.filterwarnings("ignore", category=UserWarning)
+warnings.filterwarnings("ignore", category=FutureWarning)
+plt.rcParams['font.family'] = ['Microsoft JhengHei']
+plt.rcParams['axes.unicode_minus'] = False
+
+
+# 參考 gemini 的建議，再來調整
+
+def load_data():
+    """
+    載入並對資料進行固定的預處理。
+    此函式只在應用程式啟動時執行一次。
+    """
+    df = pd.read_csv('sampled_data.csv')
+    df_processed = df.copy()
+    df_processed = df_processed.drop('id', axis=1)
+    df_processed['Gender'] = df_processed['Gender'].apply(lambda x: 1 if x == 'Male' else 0)
+    age_mapping = {'< 1 Year': 0, '1-2 Year': 1, '> 2 Years': 2}
+    df_processed['Vehicle_Age'] = df_processed['Vehicle_Age'].map(age_mapping)
+    df_processed['Vehicle_Damage'] = df_processed['Vehicle_Damage'].apply(lambda x: 1 if x == 'Yes' else 0)
+    return df, df_processed
+
+df_original, df_processed = load_data()
+ALL_FEATURES = [col for col in df_processed.columns if col != 'Response']
+NUMERICAL_FEATURES = [f for f in df_original.select_dtypes(include=np.number).columns.tolist() if f in ALL_FEATURES]
+
+# --- EDA 相關函式 ---
+def update_eda_section(selected_features):
+    if not selected_features:
+        return pd.DataFrame(), pd.DataFrame(), gr.update(choices=[], value=None), None
+    stats = df_processed[selected_features].describe().T.reset_index().rename(columns={'index': 'Feature'})
+    corrs = df_processed[selected_features + ['Response']].corr(numeric_only=True)['Response'].drop('Response').to_frame().reset_index()
+    corrs.columns = ['Feature', 'Correlation with Response']
+    first_feature_plot = generate_feature_plot(selected_features[0])
+    plot_selector_update = gr.update(choices=selected_features, value=selected_features[0])
+    return stats, corrs, plot_selector_update, first_feature_plot
+
+def generate_feature_plot(feature):
+    if not feature: return None
+    fig, ax = plt.subplots()
+    if feature in NUMERICAL_FEATURES:
+        sns.histplot(data=df_processed, x=feature, hue='Response', kde=True, ax=ax, palette='viridis', multiple="stack")
+        ax.set_title(f'"{feature}" 的直方圖 (依 Response 分色)')
+    else:
+        sns.countplot(data=df_processed, x=feature, hue='Response', ax=ax, palette='viridis')
+        ax.set_title(f'"{feature}" 的計數長條圖 (依 Response 分色)')
+    plt.tight_layout()
+    return fig
+
+# --- 核心訓練與評估函式 ---
+def train_and_evaluate(history_log, model_name, features, lr_c, lr_solver, dt_criterion, dt_max_depth, xgb_n_estimators, xgb_max_depth, xgb_learning_rate, svm_c, svm_kernel):
+    """
+    當使用者點擊 "執行模型訓練" 按鈕時觸發。
+    整合了資料準備、模型訓練、評估、結果視覺化以及紀錄日誌的完整流程。
+    """
+    if not features:
+        # 如果沒有選擇特徵，只回傳錯誤訊息和空的日誌
+        return "錯誤：請至少選擇一個特徵！", None, None, None, pd.DataFrame(history_log, columns=LOG_COLUMNS), history_log
+    
+    # --- 1. 資料準備 ---
+    X = df_processed[features]
+    y = df_processed['Response']
+    X_scaled = X.copy()
+    numerical_cols_in_x = [f for f in NUMERICAL_FEATURES if f in X_scaled.columns]
+    if numerical_cols_in_x:
+        scaler = StandardScaler()
+        X_scaled[numerical_cols_in_x] = scaler.fit_transform(X_scaled[numerical_cols_in_x])
+    X_train, X_test, y_train, y_test = train_test_split(X_scaled, y, test_size=0.3, random_state=42, stratify=y)
+
+    # --- 2. 模型選擇與訓練 ---
+    params = {}
+    if model_name == '羅吉斯回歸':
+        params = {'C': lr_c, 'solver': lr_solver}
+        X_train_sm = sm.add_constant(X_train); X_test_sm = sm.add_constant(X_test)
+        logit_model = sm.Logit(y_train, X_train_sm)
+        result = logit_model.fit(disp=0)
+        y_pred_proba = result.predict(X_test_sm); y_pred = (y_pred_proba > 0.5).astype(int)
+        importances, title = result.tvalues.drop('const', errors='ignore'), '特徵 t-值 (Feature t-values)'
+    else:
+        if model_name == '決策樹':
+            params = {'criterion': dt_criterion, 'max_depth': dt_max_depth}
+            model = DecisionTreeClassifier(**params, random_state=42, class_weight='balanced')
+        elif model_name == 'XGBoost':
+            params = {'n_estimators': int(xgb_n_estimators), 'max_depth': int(xgb_max_depth), 'learning_rate': xgb_learning_rate}
+            scale_pos_weight = y_train.value_counts()[0] / y_train.value_counts()[1]
+            model = xgb.XGBClassifier(**params, scale_pos_weight=scale_pos_weight, use_label_encoder=False, eval_metric='logloss', random_state=42)
+        elif model_name == 'SVM':
+            params = {'C': svm_c, 'kernel': svm_kernel}
+            model = SVC(**params, probability=True, random_state=42, class_weight='balanced')
+        
+        model.fit(X_train, y_train)
+        y_pred = model.predict(X_test); y_pred_proba = model.predict_proba(X_test)[:, 1]
+
+        if model_name == 'SVM' and svm_kernel == 'linear': importances, title = model.coef_[0], '特徵係數 (Feature Coefficients)'
+        elif model_name in ['決策樹', 'XGBoost']: importances, title = model.feature_importances_, '特徵重要性 (Feature Importance)'
+        else: importances, title = None, '特徵重要性'
+
+    # --- 3. 評估與繪圖 ---
+    accuracy = accuracy_score(y_test, y_pred)
+    report = classification_report(y_test, y_pred, target_names=['不感興趣 (0)', '感興趣 (1)'])
+    auc_score = f"ROC-AUC 分數: {roc_auc_score(y_test, y_pred_proba):.4f}"
+    cm = confusion_matrix(y_test, y_pred)
+    fig_cm, ax_cm = plt.subplots(); sns.heatmap(cm, annot=True, fmt='d', cmap='Blues', ax=ax_cm, xticklabels=['預測為 0', '預測為 1'], yticklabels=['實際為 0', '實際為 1']); ax_cm.set_title('混淆矩陣'); ax_cm.set_xlabel('預測標籤'); ax_cm.set_ylabel('實際標籤'); plt.tight_layout()
+
+    fig_imp, ax_imp = plt.subplots()
+    if importances is not None:
+        feature_imp = pd.Series(importances, index=features).sort_values(ascending=False)
+        sns.barplot(x=feature_imp, y=feature_imp.index, ax=ax_imp); ax_imp.set_title(title)
+    else:
+        ax_imp.text(0.5, 0.5, '此模型/核心無法直接顯示特徵重要性', ha='center', va='center'); ax_imp.set_title(title)
+    plt.tight_layout()
+
+    # --- 4. 紀錄日誌 ---
+    new_log_entry = [
+        pd.Timestamp.now().strftime('%Y-%m-%d %H:%M:%S'),
+        model_name,
+        ', '.join(features),
+        json.dumps(params),
+        f"{accuracy:.4f}"
+    ]
+    # 將新紀錄加到歷史紀錄的開頭
+    updated_log = [new_log_entry] + history_log
+    log_df = pd.DataFrame(updated_log, columns=LOG_COLUMNS)
+
+    return report, auc_score, fig_cm, fig_imp, log_df, updated_log
+
+# --- Gradio 介面設計 ---
+LOG_COLUMNS = ["時間", "模型", "特徵", "參數", "準確率"]
+
+with gr.Blocks(theme=gr.themes.Soft()) as demo:
+    # 用於儲存日誌的隱藏狀態元件
+    log_state = gr.State([])
+
+    gr.Markdown("# 互動式投保預測模型分析器")
+    gr.Markdown("在左側選擇特徵並點擊按鈕進行探索，或調整參數後點擊按鈕以訓練模型。")
+    with gr.Row():
+        with gr.Column(scale=1):
+            gr.Markdown("## 1. 特徵選擇與探索")
+            feature_selector = gr.CheckboxGroup(ALL_FEATURES, label="選擇特徵", value=['Previously_Insured', 'Vehicle_Damage', 'Policy_Sales_Channel', 'Vehicle_Age', 'Age'])
+            with gr.Row():
+                select_all_btn = gr.Button("全部選取"); deselect_all_btn = gr.Button("全部清除")
+            with gr.Accordion("特徵探索 (EDA)", open=True):
+                eda_run_btn = gr.Button("執行資料探索", variant="secondary")
+                eda_stats = gr.DataFrame(label="敘述性統計")
+                eda_corr = gr.DataFrame(label="與目標 'Response' 的相關係數")
+                eda_plot_selector = gr.Dropdown(label="選擇要視覺化的特徵")
+                eda_plot = gr.Plot(label="視覺化")
+            gr.Markdown("## 2. 模型選擇與超參數調整")
+            model_selector = gr.Dropdown(['羅吉斯回歸', '決策樹', 'XGBoost', 'SVM'], label="選擇模型", value='決策樹')
+            with gr.Group(visible=False) as lr_box:
+                gr.Markdown("#### 羅吉斯回歸"); lr_c = gr.Slider(0.01, 10.0, value=1.0, step=0.01, label="C (正規化強度, statsmodels中未使用)"); lr_solver = gr.Dropdown(['lbfgs', 'liblinear', 'saga'], value='lbfgs', label="優化演算法 (statsmodels中未使用)")
+            with gr.Group(visible=True) as dt_box:
+                gr.Markdown("#### 決策樹"); dt_criterion = gr.Radio(['gini', 'entropy'], value='gini', label="評估標準"); dt_max_depth = gr.Slider(3, 30, value=8, step=1, label="最大深度")
+            with gr.Group(visible=False) as xgb_box:
+                gr.Markdown("#### XGBoost"); xgb_n_estimators = gr.Slider(50, 500, value=100, step=10, label="樹的數量"); xgb_max_depth = gr.Slider(3, 15, value=5, step=1, label="最大深度"); xgb_learning_rate = gr.Slider(0.01, 0.3, value=0.1, step=0.01, label="學習率")
+            with gr.Group(visible=False) as svm_box:
+                gr.Markdown("#### SVM"); svm_c = gr.Slider(0.01, 10.0, value=1.0, step=0.01, label="C (懲罰參數)"); svm_kernel = gr.Radio(['linear', 'rbf', 'poly'], value='linear', label="核心")
+            run_btn = gr.Button("🚀 執行模型訓練", variant="primary")
+        with gr.Column(scale=2):
+            gr.Markdown("## 3. 模型評估��果")
+            model_output_report = gr.Textbox(label="分類報告", lines=10)
+            model_output_auc = gr.Textbox(label="AUC 分數")
+            model_plot_cm = gr.Plot(label="混淆矩陣")
+            model_plot_importance = gr.Plot(label="特徵重要性/係數")
+    
+    with gr.Accordion("操作紀錄 (History Log)", open=False):
+        log_df_display = gr.DataFrame(headers=LOG_COLUMNS, datatype=["str", "str", "str", "str", "str"])
+
+    # --- 事件處理 ---
+    eda_run_btn.click(update_eda_section, inputs=feature_selector, outputs=[eda_stats, eda_corr, eda_plot_selector, eda_plot])
+    eda_plot_selector.change(generate_feature_plot, inputs=eda_plot_selector, outputs=eda_plot)
+    def show_hyperparameters(model_name): return {lr_box: gr.update(visible=model_name == '羅吉斯回歸'), dt_box: gr.update(visible=model_name == '決策樹'), xgb_box: gr.update(visible=model_name == 'XGBoost'), svm_box: gr.update(visible=model_name == 'SVM')}
+    model_selector.change(show_hyperparameters, inputs=model_selector, outputs=[lr_box, dt_box, xgb_box, svm_box])
+    def select_all_features(): return gr.update(value=ALL_FEATURES)
+    def deselect_all_features(): return gr.update(value=[])
+    select_all_btn.click(select_all_features, None, feature_selector)
+    deselect_all_btn.click(deselect_all_features, None, feature_selector)
+    
+    run_btn.click(
+        train_and_evaluate, 
+        inputs=[log_state, model_selector, feature_selector, lr_c, lr_solver, dt_criterion, dt_max_depth, xgb_n_estimators, xgb_max_depth, xgb_learning_rate, svm_c, svm_kernel], 
+        outputs=[model_output_report, model_output_auc, model_plot_cm, model_plot_importance, log_df_display, log_state]
+    )
+
+if __name__ == "__main__":
+    demo.launch()

requirements.txt

@@ -0,0 +1,8 @@
+gradio
+pandas
+numpy
+matplotlib
+seaborn
+scikit-learn
+xgboost
+statsmodels