在數(shù)據(jù)預(yù)處理中，我們從Excel文件中讀取數(shù)據(jù)，并劃分特征和目標(biāo)變量，與往期代碼不同的是，這次我們沒有單獨劃分驗證集，因為K折交叉驗證會在輸入的訓(xùn)練集中自動進(jìn)行驗證劃分，這樣，整個訓(xùn)練集都被充分利用，提高了模型評估的可靠性，同時我們的測試集保持不變，用于最終的模型評估

模型定義

# CatBoost模型參數(shù)

params_cat = {

    'learning_rate': 0.02,

    'iterations': 1000,

    'depth': 6,

    'eval_metric': 'RMSE',

    'random_seed': 42,

    'verbose': 100

}



# NGBoost模型參數(shù)

params_ngb = {

    'learning_rate': 0.02,

    'n_estimators': 1000,

    'verbose': False,

    'random_state': 42,

    'natural_gradient': True

}



# 定義平均模型類

class AverageModel:

    def __init__(self, models):

        self.models = models



    def fit(self, X, y, X_val, y_val):

        for model in self.models:

            if isinstance(model, NGBRegressor):

                model.fit(X, y, X_val=X_val, Y_val=y_val)  # NGBoost的fit方法接受驗證集參數(shù)

            elif isinstance(model, CatBoostRegressor):

                model.fit(X, y, eval_set=(X_val, y_val), use_best_model=True, verbose=False)



    def predict(self, X):

        predictions = []

        for model in self.models:

            predictions.append(model.predict(X))

        return sum(predictions) / len(predictions)

定義CatBoost和NGBoost模型的參數(shù)，并通過一個名為AverageModel的類來訓(xùn)練這些模型，該類在訓(xùn)練時根據(jù)不同模型調(diào)用相應(yīng)的fit方法并在預(yù)測時平均所有模型的預(yù)測結(jié)果，當(dāng)然也可以改進(jìn)組合策略在預(yù)測時通過不同的策略（如加權(quán)平均、投票等）綜合所有模型的預(yù)測結(jié)果

k折交叉驗證下訓(xùn)練模型

# 定義k折交叉驗證

kf = KFold(n_splits=5, shuffle=True, random_state=42)

scores = []

best_score = float('inf')

best_model = None



# 進(jìn)行k折交叉驗證

for fold, (train_index, val_index) in enumerate(kf.split(X_train, y_train)):

    X_train_fold, X_val_fold = X_train.iloc[train_index], X_train.iloc[val_index]

    y_train_fold, y_val_fold = y_train.iloc[train_index], y_train.iloc[val_index]



    # 計算當(dāng)前訓(xùn)練集的均值和標(biāo)準(zhǔn)差

    x_mean_fold = X_train_fold.mean()

    x_std_fold = X_train_fold.std()

    y_mean_fold = y_train_fold.mean()

    y_std_fold = y_train_fold.std()



    # 標(biāo)準(zhǔn)化

    X_train_fold = (X_train_fold - x_mean_fold) / x_std_fold

    y_train_fold = (y_train_fold - y_mean_fold) / y_std_fold

    X_val_fold = (X_val_fold - x_mean_fold) / x_std_fold

    y_val_fold = (y_val_fold - y_mean_fold) / y_std_fold



    # 創(chuàng)建和訓(xùn)練組合模型

    current_model = AverageModel([NGBRegressor(**params_ngb), CatBoostRegressor(**params_cat)])

    current_model.fit(X_train_fold, y_train_fold, X_val_fold, y_val_fold)



    # 預(yù)測并計算得分

    y_val_pred = current_model.predict(X_val_fold)

    score = mean_squared_error(y_val_fold * y_std_fold + y_mean_fold, y_val_pred * y_std_fold + y_mean_fold, squared=False)  # 反標(biāo)準(zhǔn)化得分



    scores.append(score)

    print(f'Fold {fold + 1} RMSE: {score}')



    # 保留得分最好的模型

    if score < best_score:

        best_score = score

        best_model = current_model



print(f'Best RMSE: {best_score}')

定義了5折交叉驗證，通過對每折的訓(xùn)練集和驗證集進(jìn)行標(biāo)準(zhǔn)化后訓(xùn)練NGBRegressor和CatBoostRegressor組合模型，并在每折計算驗證集的RMSE來評估模型性能，最終選出得分最好的模型作為最佳模型

模型預(yù)測

# 使用最佳模型進(jìn)行測試集預(yù)測

x_mean_final = X_train.mean()

x_std_final = X_train.std()

X_test_standardized = (X_test - x_mean_final) / x_std_final

y_test_standardized = (y_test - y.mean()) / y.std()

y_test_pred_standardized = best_model.predict(X_test_standardized)

y_test_pred = y_test_pred_standardized * y.std() + y.mean()

# 輸出反標(biāo)準(zhǔn)化后的預(yù)測結(jié)果

results = pd.DataFrame({'Actual': y_test, 'Predicted': y_test_pred})

使用訓(xùn)練集的均值和標(biāo)準(zhǔn)差對測試集進(jìn)行標(biāo)準(zhǔn)化，并用最佳模型預(yù)測測試集的目標(biāo)值，再將預(yù)測結(jié)果反標(biāo)準(zhǔn)化以輸出實際值和預(yù)測值的比較

詳細(xì)評價指標(biāo)

from sklearn import metrics

import numpy as np

y_pred_list = y_test_pred.tolist()  # 或者 y_pred_array = np.array(y_pred)

mse = metrics.mean_squared_error(y_test, y_pred_list)

rmse = np.sqrt(mse)

mae = metrics.mean_absolute_error(y_test, y_pred_list)

r2 = metrics.r2_score(y_test, y_pred_list)

print("均方誤差 (MSE):", mse)

print("均方根誤差 (RMSE):", rmse)

print("平均絕對誤差 (MAE):", mae)

print("擬合優(yōu)度 (R-squared):", r2)

可視化對比

# 可視化 results 數(shù)據(jù)框

plt.figure(figsize=(10, 6), dpi=300)

plt.scatter(results['Predicted'], results['Actual'], color='blue', edgecolor='k', s=50, alpha=0.6, label='預(yù)測值 vs 真實值')

plt.title('預(yù)測值與真實值對比圖', fontsize=16)

plt.xlabel('預(yù)測值', fontsize=14)

plt.ylabel('真實值', fontsize=14)

max_val = max(results.max())

min_val = min(results.min())

plt.plot([min_val, max_val], [min_val, max_val], color='red', linestyle='--', linewidth=2, label='x=y')

plt.grid(True, linestyle='--', alpha=0.7)

plt.legend()

plt.show()

本文章轉(zhuǎn)載微信公眾號@Python機(jī)器學(xué)習(xí)AI