保险保费预测：左偏数据处理优化求助（Yeo-Johnson效果不佳）

我目前负责一项保险保费预测任务，特征包括职业描述、雇员数量、营业额等，正在对比K近邻（K-Nearest Neighbours）和XGBoost回归器的效果。

当前数据呈现左偏分布，以下是数据样本：

我尝试用Yeo-Johnson方法将数据转换为接近正态分布（该方法适用于偏态数据），但模型得分没有提升，希望得到处理偏态数据的建议。以下是我使用XGBoost的代码：

Data = pd.DataFrame(Data)

Data = Data.fillna(0)

Data['Quotation Status Description'] = Data["Quotation Status Description"].map({'Bound':1, "Quoted":0, "Declined":2})

Data = pd.get_dummies(Data, columns=["Job role"])

X = Data.drop('Quotation Premium Amount', axis=1)
y= Data['Quotation Premium Amount']

pt = PowerTransformer(method="yeo-johnson", standardize=True)

X_train, X_test, y_train_raw, y_test_raw = train_test_split(X,y, test_size=0.2)

y_train = pt.fit_transform(y_train_raw.values.reshape(-1, 1))
y_test = pt.transform(y_test_raw.values.reshape(-1, 1))

model=xgb.XGBRegressor(objective="reg:squarederror", eval_metric="rmse")

#define the hyperparameters
params = {"max_depth": [3, 5, 10, 20],
          "n_estimators": [250, 500,700, 1000],
          "learning_rate":[0.01, 0.015]}

#grid search
gridsearch = GridSearchCV(model, params, cv=5,scoring='neg_root_mean_squared_error', verbose=1)
gridsearch.fit(X_train, y_train)

#Best model
best_model=gridsearch.best_estimator_
y_preds =  best_model.predict(X_test)

y_pred_orig = pt.inverse_transform(y_preds.reshape(-1, 1))
y_test_orig = pt.inverse_transform(y_test.reshape(-1, 1))

#Best model
best_model=gridsearch.best_estimator_

print("R² (original):", r2_score(y_test_orig, y_pred_orig))
print("MSE:", mean_squared_error(y_test_orig, y_pred_orig))
print("RMSE:", np.sqrt(mean_squared_error(y_test_orig, y_pred_orig)))

print("Best Params: ",gridsearch.best_params_)

内容的提问来源于stack exchange，提问作者Red_bull