模型训练

GPT 4o-Canvas：数据分析再次发生变革！

2024 年 10 月 17 日

三天前，OpenAI发布了GPT 4o的新功能——Canvas。如果你熟悉Claude中的Artifact工具，这个功能和它类似。

使用这个功能可以简化数据分析。在本文中，我们将一起探索如何做到这一点。为此，我们将使用来自Kaggle的全球地震数据集。

全球地震数据

https://www.kaggle.com/datasets/shreyasur965/recent-earthquakes

这个数据集有43个属性，包括震级、位置、深度和地震学测量，详细记录了全球1137次地震的信息。

GPT 4o-Canvas

首先，从可用的模型列表中选择这个模型。

好了，现在让我们使用这个提示结构，它将帮助你更明智地分析数据集。但在此之前，先上传你从Kaggle下载的数据集。

Here is the dataset information
[Paste Dataset information from Kaggle]
Write me a machine learning code that predicts magnitude of earthquakes.

这里是代码。

但是你注意到了吗？它创建了自己的数据集，而不是使用我们上传的那个。

这没关系；大语言模型偶尔会出错，这就是它们仍然需要人类的原因。

这里是调整提示词的方法。

You should not create an example data, we already have the data that I sent to you an
look from there and write code here accordingly

好了。如你所见，它使用了我们的数据集，但这次它自己选择了特征。我们需要对此进行更改。

这里是我使用的提示词。

you should select features based on the feature importance, maybe gini index?

现在，让我们向ChatGPT输入信息。

Here all features.
[paste df.info codes output here- all column names]
You should select all at first, and then label encoding then and 
select important ones with gini

在这里，我意识到我们使用了很多不同的方法来创建最佳模型，比如：

归一化
应用5种不同的机器学习模型
通过Gini Importance选择不同的模型
对数变换
降维

我让GPT 4o-Canvas为这些方法生成代码。这里是我最后的代码。

import pandas as pd
from sklearn.model_selection import train_test_split
from sklearn.ensemble import RandomForestRegressor, GradientBoostingRegressor
from sklearn.linear_model import LinearRegression
from sklearn.tree import DecisionTreeRegressor
from sklearn.svm import SVR
from sklearn.metrics import mean_squared_error, r2_score
from sklearn.preprocessing import LabelEncoder, StandardScaler, MinMaxScaler
from sklearn.decomposition import PCA
import numpy as np
import matplotlib.pyplot as plt
# Load the dataset
df = pd.read_csv('/mnt/data/earthquakes.csv')
# Select all features initially
all_features = df.columns.tolist()
all_features.remove('magnitude')# Remove target variable from features
# Handle categorical features using Label Encoding
categorical_features = df.select_dtypes(include=['object']).columns
label_encoders = {}
for feature in categorical_features:
le = LabelEncoder()
df[feature] = le.fit_transform(df[feature].astype(str))
label_encoders[feature] = le
# Drop rows with missing values in the selected features and target
df = df.dropna(subset=all_features + ['magnitude'])
# Define the input features (X) and the target variable (y)
X = df[all_features]
y = df['magnitude']
# Check for skewness and apply log transformation where necessary
for col in X.columns:
if X[col].skew() > 1 or X[col].skew() < -1:
X.loc[:, col] = np.log1p(X[col])
# Normalize the features using StandardScaler
scaler_standard = StandardScaler()
X_standard = scaler_standard.fit_transform(X)
# Normalize the features using MinMaxScaler
scaler_minmax = MinMaxScaler()
X_minmax = scaler_minmax.fit_transform(X)
# Dimensionality Reduction using PCA (if number of features is high)
pca_threshold = 0.95
if X_standard.shape[1] > 10:
pca_standard = PCA(n_components=pca_threshold)
X_standard = pca_standard.fit_transform(X_standard)
pca_minmax = PCA(n_components=pca_threshold)
X_minmax = pca_minmax.fit_transform(X_minmax)
# Split the data into training and testing sets for both normalization methods
X_train_standard, X_test_standard, y_train, y_test = train_test_split(X_standard, y, test_size=0.2, random_state=42)
X_train_minmax, X_test_minmax, _, _ = train_test_split(X_minmax, y, test_size=0.2, random_state=42)
# Initialize RandomForestRegressor to determine feature importance
feature_importance_model = RandomForestRegressor(n_estimators=100, random_state=42)
feature_importance_model.fit(X_train_standard, y_train)
# Calculate feature importance using the features used in the training set after transformations
actual_features = X.columns[:X_train_standard.shape[1]]
feature_importances = feature_importance_model.feature_importances_
feature_importance_df = pd.DataFrame({'feature': actual_features, 'importance': feature_importances})
feature_importance_df = feature_importance_df.sort_values(by='importance', ascending=False)
# Select the most important features (top 8 and top 10)
top_8_features = feature_importance_df['feature'].head(8).tolist()
top_10_features = feature_importance_df['feature'].head(10).tolist()
# Define input features for top 8 and top 10 feature sets
X_top_8 = df[top_8_features].copy()
X_top_10 = df[top_10_features].copy()
# Check for skewness in top features and apply log transformation where necessary
for col in X_top_8.columns:
if X_top_8[col].skew() > 1 or X_top_8[col].skew() < -1:
X_top_8.loc[:, col] = np.log1p(X_top_8[col])
for col in X_top_10.columns:
if X_top_10[col].skew() > 1 or X_top_10[col].skew() < -1:
X_top_10.loc[:, col] = np.log1p(X_top_10[col])
# Normalize the selected top features using both scalers
X_top_8_standard = scaler_standard.fit_transform(X_top_8)
X_top_8_minmax = scaler_minmax.fit_transform(X_top_8)
X_top_10_standard = scaler_standard.fit_transform(X_top_10)
X_top_10_minmax = scaler_minmax.fit_transform(X_top_10)
# Dimensionality Reduction using PCA for top 8 and top 10 features
if X_top_8_standard.shape[1] > 5:
pca_top_8_standard = PCA(n_components=pca_threshold)
X_top_8_standard = pca_top_8_standard.fit_transform(X_top_8_standard)
pca_top_8_minmax = PCA(n_components=pca_threshold)
X_top_8_minmax = pca_top_8_minmax.fit_transform(X_top_8_minmax)
if X_top_10_standard.shape[1] > 5:
pca_top_10_standard = PCA(n_components=pca_threshold)
X_top_10_standard = pca_top_10_standard.fit_transform(X_top_10_standard)
pca_top_10_minmax = PCA(n_components=pca_threshold)
X_top_10_minmax = pca_top_10_minmax.fit_transform(X_top_10_minmax)
# Split the data for top 8 and top 10 feature sets
X_train_8_standard, X_test_8_standard, _, _ = train_test_split(X_top_8_standard, y, test_size=0.2, random_state=42)
X_train_8_minmax, X_test_8_minmax, _, _ = train_test_split(X_top_8_minmax, y, test_size=0.2, random_state=42)
X_train_10_standard, X_test_10_standard, _, _ = train_test_split(X_top_10_standard, y, test_size=0.2, random_state=42)
X_train_10_minmax, X_test_10_minmax, _, _ = train_test_split(X_top_10_minmax, y, test_size=0.2, random_state=42)
# Define the models to be used
models = {
'Random Forest': RandomForestRegressor(n_estimators=100, random_state=42),
'Gradient Boosting': GradientBoostingRegressor(n_estimators=100, random_state=42),
'Linear Regression': LinearRegression(),
'Decision Tree': DecisionTreeRegressor(random_state=42),
'Support Vector Regressor': SVR()
}
# Train each model and evaluate performance for each combination
model_results = {}
normalizations = [
('StandardScaler - Top 8 Features', X_train_8_standard, X_test_8_standard),
('MinMaxScaler - Top 8 Features', X_train_8_minmax, X_test_8_minmax),
('StandardScaler - Top 10 Features', X_train_10_standard, X_test_10_standard),
('MinMaxScaler - Top 10 Features', X_train_10_minmax, X_test_10_minmax)
]
for normalization_name, X_train, X_test in normalizations:
for model_name, model in models.items():
model.fit(X_train, y_train)
y_pred = model.predict(X_test)
mse = mean_squared_error(y_test, y_pred)
r2 = r2_score(y_test, y_pred)
model_results[f'{model_name} ({normalization_name})'] = {'MSE': mse, 'R2': r2}
print(f"{model_name} ({normalization_name}) - Mean Squared Error: {mse}, R^2 Score: {r2}")
# Plot the results
model_names = list(model_results.keys())
mse_values = [model_results[name]['MSE'] for name in model_names]
r2_values = [model_results[name]['R2'] for name in model_names]
fig, ax = plt.subplots(1, 2, figsize=(20, 10))
# Bar plot for MSE
ax[0].barh(model_names, mse_values, color='skyblue')
ax[0].set_title('Mean Squared Error Comparison')
ax[0].set_xlabel('MSE')
# Bar plot for R2 Score
ax[1].barh(model_names, r2_values, color='lightgreen')
ax[1].set_title('R^2 Score Comparison')
ax[1].set_xlabel('R^2 Score')
plt.tight_layout()
plt.show()

这里是输出的第一部分。

Random Forest (StandardScaler - Top 8 Features) - Mean Squared Error: 0.036684462750000014, R^2 Score: 0.8779777294037107
Gradient Boosting (StandardScaler - Top 8 Features) - Mean Squared Error: 0.03171284438291409, R^2 Score: 0.8945146531096477
Linear Regression (StandardScaler - Top 8 Features) - Mean Squared Error: 0.2030328908150269, R^2 Score: 0.3246586569411304
Decision Tree (StandardScaler - Top 8 Features) - Mean Squared Error: 0.014955, R^2 Score: 0.9502556962820041
Support Vector Regressor (StandardScaler - Top 8 Features) - Mean Squared Error: 0.19093007009631902, R^2 Score: 0.3649158545122343
Random Forest (MinMaxScaler - Top 8 Features) - Mean Squared Error: 0.0368944044999999, R^2 Score: 0.8772794073592385
Gradient Boosting (MinMaxScaler - Top 8 Features) - Mean Squared Error: 0.02585888543990656, R^2 Score: 0.9139864760192863
Linear Regression (MinMaxScaler - Top 8 Features) - Mean Squared Error: 0.2030123102574625, R^2 Score: 0.3247271133440839
Decision Tree (MinMaxScaler - Top 8 Features) - Mean Squared Error: 0.021312499999999988, R^2 Score: 0.9291089620200744
Support Vector Regressor (MinMaxScaler - Top 8 Features) - Mean Squared Error: 0.1900837276420734, R^2 Score: 0.3677310143981205
Random Forest (StandardScaler - Top 10 Features) - Mean Squared Error: 0.014616167999999801, R^2 Score: 0.9513827415456206
Gradient Boosting (StandardScaler - Top 10 Features) - Mean Squared Error: 0.005215391903357051, R^2 Score: 0.9826522207389522
Linear Regression (StandardScaler - Top 10 Features) - Mean Squared Error: 0.024659998640491166, R^2 Score: 0.9179742920723531
Decision Tree (StandardScaler - Top 10 Features) - Mean Squared Error: 0.01719999999999999, R^2 Score: 0.9427882297593093
Support Vector Regressor (StandardScaler - Top 10 Features) - Mean Squared Error: 0.012001988809648145, R^2 Score: 0.9600781961506436
Random Forest (MinMaxScaler - Top 10 Features) - Mean Squared Error: 0.013746897999999952, R^2 Score: 0.9542741645408019
Gradient Boosting (MinMaxScaler - Top 10 Features) - Mean Squared Error: 0.0066618489718651116, R^2 Score: 0.9778409201885739
Linear Regression (MinMaxScaler - Top 10 Features) - Mean Squared Error: 0.024620634208737134, R^2 Score: 0.918105228631956
Decision Tree (MinMaxScaler - Top 10 Features) - Mean Squared Error: 0.02131249999999998, R^2 Score: 0.9291089620200744
Support Vector Regressor (MinMaxScaler - Top 10 Features) - Mean Squared Error: 0.01063026736761726, R^2 Score: 0.9646409061492308