513 KiB
513 KiB
In [ ]:
import pandas as pd
import matplotlib.pyplot as plt
df = pd.read_csv("C://Users//annal//aim//static//csv//Forbes_Billionaires.csv")
print(df.columns)
Подготовим данные: категоризируем колонку age¶
In [2]:
print(df.isnull().sum())
print()
# Есть ли пустые значения признаков
print(df.isnull().any())
print()
# Процент пустых значений признаков
for i in df.columns:
null_rate = df[i].isnull().sum() / len(df) * 100
if null_rate > 0:
print(f"{i} процент пустых значений: %{null_rate:.2f}")
In [2]:
bins = [0, 30, 40, 50, 60, 70, 80, 101] # границы для возрастных категорий
labels = ['Under 30', '30-40', '40-50', '50-60', '60-70', '70-80', '80+'] # метки для категорий
df["Age_category"] = pd.cut(df['Age'], bins=bins, labels=labels, right=False)
# Удаляем оригинальные колонки 'country', 'industry' и 'source' из исходного DataFrame
df.drop(columns=['Age'], inplace=True)
# Просмотр результата
print(df.head())
In [27]:
from utils import split_stratified_into_train_val_test
X_train, X_val, X_test, y_train, y_val, y_test = split_stratified_into_train_val_test(
df, stratify_colname="Age_category", frac_train=0.80, frac_val=0, frac_test=0.20, random_state=9
)
display("X_train", X_train)
display("y_train", y_train)
display("X_test", X_test)
display("y_test", y_test)
Формирование конвейера для классификации данных¶
preprocessing_num -- конвейер для обработки числовых данных: заполнение пропущенных значений и стандартизация¶
preprocessing_cat -- конвейер для обработки категориальных данных: заполнение пропущенных данных и унитарное кодирование¶
features_preprocessing -- трансформер для предобработки признаков¶
features_engineering -- трансформер для конструирования признаков¶
drop_columns -- трансформер для удаления колонок¶
pipeline_end -- основной конвейер предобработки данных и конструирования признаков¶
In [37]:
from sklearn.compose import ColumnTransformer
from sklearn.preprocessing import StandardScaler, OneHotEncoder
from sklearn.impute import SimpleImputer
from sklearn.pipeline import Pipeline
import pandas as pd
# Исправляем ColumnTransformer с сохранением имен колонок
columns_to_drop = ["Age_category", "Rank ", "Name"]
num_columns = [
column
for column in X_train.columns
if column not in columns_to_drop and X_train[column].dtype != "object"
]
cat_columns = [
column
for column in X_train.columns
if column not in columns_to_drop and X_train[column].dtype == "object"
]
# Предобработка числовых данных
num_imputer = SimpleImputer(strategy="median")
num_scaler = StandardScaler()
preprocessing_num = Pipeline(
[
("imputer", num_imputer),
("scaler", num_scaler),
]
)
# Предобработка категориальных данных
cat_imputer = SimpleImputer(strategy="constant", fill_value="unknown")
cat_encoder = OneHotEncoder(handle_unknown="ignore", sparse_output=False, drop="first")
preprocessing_cat = Pipeline(
[
("imputer", cat_imputer),
("encoder", cat_encoder),
]
)
# Общая предобработка признаков
features_preprocessing = ColumnTransformer(
verbose_feature_names_out=True, # Сохраняем имена колонок
transformers=[
("prepocessing_num", preprocessing_num, num_columns),
("prepocessing_cat", preprocessing_cat, cat_columns),
],
remainder="drop" # Убираем неиспользуемые столбцы
)
# Итоговый конвейер
pipeline_end = Pipeline(
[
("features_preprocessing", features_preprocessing),
]
)
# Преобразуем данные
preprocessing_result = pipeline_end.fit_transform(X_train)
# Создаем DataFrame с правильными именами колонок
preprocessed_df = pd.DataFrame(
preprocessing_result,
columns=pipeline_end.get_feature_names_out(),
index=X_train.index, # Сохраняем индексы
)
preprocessed_df
Out[37]:
Формирование набора моделей для классификации¶
logistic -- логистическая регрессия¶
ridge -- гребневая регрессия¶
decision_tree -- дерево решений¶
knn -- k-ближайших соседей¶
naive_bayes -- наивный Байесовский классификатор¶
gradient_boosting -- метод градиентного бустинга (набор деревьев решений)¶
random_forest -- метод случайного леса (набор деревьев решений)¶
mlp -- многослойный персептрон (нейронная сеть)¶
In [38]:
from sklearn import ensemble, linear_model, naive_bayes, neighbors, neural_network, tree
class_models = {
"logistic": {"model": linear_model.LogisticRegression()},
# "ridge": {"model": linear_model.RidgeClassifierCV(cv=5, class_weight="balanced")},
"ridge": {"model": linear_model.LogisticRegression(penalty="l2", class_weight="balanced")},
"decision_tree": {
"model": tree.DecisionTreeClassifier(max_depth=7, random_state=9)
},
"knn": {"model": neighbors.KNeighborsClassifier(n_neighbors=7)},
"naive_bayes": {"model": naive_bayes.GaussianNB()},
"gradient_boosting": {
"model": ensemble.GradientBoostingClassifier(n_estimators=210)
},
"random_forest": {
"model": ensemble.RandomForestClassifier(
max_depth=11, class_weight="balanced", random_state=9
)
},
"mlp": {
"model": neural_network.MLPClassifier(
hidden_layer_sizes=(7,),
max_iter=500,
early_stopping=True,
random_state=9,
)
},
}
Обучение моделей на обучающем наборе данных и оценка на тестовом¶
In [43]:
y_train['Age_category'] = y_train['Age_category'].cat.codes
y_test['Age_category'] = y_test['Age_category'].cat.codes
In [44]:
import numpy as np
from sklearn import metrics
for model_name in class_models.keys():
print(f"Model: {model_name}")
model = class_models[model_name]["model"]
model_pipeline = Pipeline([("pipeline", pipeline_end), ("model", model)])
model_pipeline = model_pipeline.fit(X_train, y_train)
y_train_predict = model_pipeline.predict(X_train)
y_test_probs = model_pipeline.predict_proba(X_test)
y_test_predict = np.argmax(y_test_probs, axis=1)
class_models[model_name]["pipeline"] = model_pipeline
class_models[model_name]["probs"] = y_test_probs
class_models[model_name]["preds"] = y_test_predict
# Метрики
class_models[model_name]["Precision_train"] = metrics.precision_score(
y_train, y_train_predict, average="macro"
)
class_models[model_name]["Precision_test"] = metrics.precision_score(
y_test, y_test_predict, average="macro"
)
class_models[model_name]["Recall_train"] = metrics.recall_score(
y_train, y_train_predict, average="macro"
)
class_models[model_name]["Recall_test"] = metrics.recall_score(
y_test, y_test_predict, average="macro"
)
class_models[model_name]["Accuracy_train"] = metrics.accuracy_score(
y_train, y_train_predict
)
class_models[model_name]["Accuracy_test"] = metrics.accuracy_score(
y_test, y_test_predict
)
class_models[model_name]["ROC_AUC_test"] = metrics.roc_auc_score(
y_test, y_test_probs, multi_class="ovr"
)
class_models[model_name]["F1_train"] = metrics.f1_score(
y_train, y_train_predict, average="macro"
)
class_models[model_name]["F1_test"] = metrics.f1_score(
y_test, y_test_predict, average="macro"
)
class_models[model_name]["MCC_test"] = metrics.matthews_corrcoef(
y_test, y_test_predict
)
class_models[model_name]["Cohen_kappa_test"] = metrics.cohen_kappa_score(
y_test, y_test_predict
)
class_models[model_name]["Confusion_matrix"] = metrics.confusion_matrix(
y_test, y_test_predict
)
Сводная таблица оценок качества для использованных моделей классификации¶
In [49]:
from sklearn.metrics import ConfusionMatrixDisplay
import matplotlib.pyplot as plt
_, ax = plt.subplots(int(len(class_models) / 2), 2, figsize=(17, 17), sharex=False, sharey=False)
for index, key in enumerate(class_models.keys()):
c_matrix = class_models[key]["Confusion_matrix"]
disp = ConfusionMatrixDisplay(
confusion_matrix=c_matrix, display_labels=["Under 30", "30-40", "40-50", "50-60", "60-70", "70-80", "80+"]
).plot(ax=ax.flat[index])
disp.ax_.set_title(key)
plt.subplots_adjust(top=1, bottom=0, hspace=0.4, wspace=0.1)
plt.show()
Точность, полнота, верность (аккуратность), F-мера¶
In [52]:
class_metrics = pd.DataFrame.from_dict(class_models, "index")[
[
"Precision_train",
"Precision_test",
"Recall_train",
"Recall_test",
"Accuracy_train",
"Accuracy_test",
"F1_train",
"F1_test",
]
]
class_metrics.sort_values(
by="Accuracy_test", ascending=False
).style.background_gradient(
cmap="plasma",
low=0.3,
high=1,
subset=["Accuracy_train", "Accuracy_test", "F1_train", "F1_test"],
).background_gradient(
cmap="viridis",
low=1,
high=0.3,
subset=[
"Precision_train",
"Precision_test",
"Recall_train",
"Recall_test",
],
)
Out[52]:
значения далеки от идела, датасет так себе...¶
In [51]:
pip install Jinja2
ROC-кривая, каппа Коэна, коэффициент корреляции Мэтьюса¶
In [53]:
class_metrics = pd.DataFrame.from_dict(class_models, "index")[
[
"Accuracy_test",
"F1_test",
"ROC_AUC_test",
"Cohen_kappa_test",
"MCC_test",
]
]
class_metrics.sort_values(by="ROC_AUC_test", ascending=False).style.background_gradient(
cmap="plasma",
low=0.3,
high=1,
subset=[
"ROC_AUC_test",
"MCC_test",
"Cohen_kappa_test",
],
).background_gradient(
cmap="viridis",
low=1,
high=0.3,
subset=[
"Accuracy_test",
"F1_test",
],
)
Out[53]:
In [54]:
best_model = str(class_metrics.sort_values(by="MCC_test", ascending=False).iloc[0].name)
display(best_model)
Вывод данных с ошибкой предсказания для оценки¶
In [56]:
preprocessing_result = pipeline_end.transform(X_test)
preprocessed_df = pd.DataFrame(
preprocessing_result,
columns=pipeline_end.get_feature_names_out(),
)
y_pred = class_models[best_model]["preds"]
error_index = y_test[y_test["Age_category"] != y_pred].index.tolist()
display(f"Error items count: {len(error_index)}")
error_predicted = pd.Series(y_pred, index=y_test.index).loc[error_index]
error_df = X_test.loc[error_index].copy()
error_df.insert(loc=1, column="Predicted", value=error_predicted)
error_df.sort_index()
Out[56]:
Многовато...¶
Пример использования обученной модели (конвейера) для предсказания¶
In [57]:
model = class_models[best_model]["pipeline"]
example_id = 450
test = pd.DataFrame(X_test.loc[example_id, :]).T
test_preprocessed = pd.DataFrame(preprocessed_df.loc[example_id, :]).T
display(test)
display(test_preprocessed)
result_proba = model.predict_proba(test)[0]
result = model.predict(test)[0]
real = int(y_test.loc[example_id].values[0])
display(f"predicted: {result} (proba: {result_proba})")
display(f"real: {real}")
Подбор гиперпараметров методом поиска по сетке¶
In [ ]:
from sklearn.model_selection import GridSearchCV
optimized_model_type = "random_forest"
random_forest_model = class_models[optimized_model_type]["pipeline"]
param_grid = {
"model__n_estimators": [10, 20, 30, 40, 50, 100, 150, 200, 250, 500],
"model__max_features": ["sqrt", "log2", 2],
"model__max_depth": [2, 3, 4, 5, 6, 7, 8, 9 ,10],
"model__criterion": ["gini", "entropy", "log_loss"],
}
gs_optomizer = GridSearchCV(
estimator=random_forest_model, param_grid=param_grid, n_jobs=-1
)
gs_optomizer.fit(X_train, y_train.values.ravel())
gs_optomizer.best_params_
Out[ ]:
Обучение модели с новыми гиперпараметрами¶
In [69]:
optimized_model = ensemble.RandomForestClassifier(
random_state=9,
criterion="gini",
max_depth=10,
max_features=2,
n_estimators=250,
)
result = {}
result["pipeline"] = Pipeline([("pipeline", pipeline_end), ("model", optimized_model)]).fit(X_train, y_train.values.ravel())
result["train_preds"] = result["pipeline"].predict(X_train)
result["probs"] = result["pipeline"].predict_proba(X_test)
result["preds"] = np.argmax(y_test_probs, axis=1)
result["Precision_train"] = metrics.precision_score(y_train, result["train_preds"],average="macro")
result["Precision_test"] = metrics.precision_score(y_test, result["preds"], average="macro")
result["Recall_train"] = metrics.recall_score(y_train, result["train_preds"], average="macro")
result["Recall_test"] = metrics.recall_score(y_test, result["preds"], average="macro")
result["Accuracy_train"] = metrics.accuracy_score(y_train, result["train_preds"])
result["Accuracy_test"] = metrics.accuracy_score(y_test, result["preds"])
result["ROC_AUC_test"] = metrics.roc_auc_score(y_test, result["probs"], multi_class="ovr")
result["F1_train"] = metrics.f1_score(y_train, result["train_preds"], average="macro")
result["F1_test"] = metrics.f1_score(y_test, result["preds"], average="macro")
result["MCC_test"] = metrics.matthews_corrcoef(y_test, result["preds"])
result["Cohen_kappa_test"] = metrics.cohen_kappa_score(y_test, result["preds"])
result["Confusion_matrix"] = metrics.confusion_matrix(y_test, result["preds"])
Формирование данных для оценки старой и новой версии модели¶
In [70]:
optimized_metrics = pd.DataFrame(columns=list(result.keys()))
optimized_metrics.loc[len(optimized_metrics)] = pd.Series(
data=class_models[optimized_model_type]
)
optimized_metrics.loc[len(optimized_metrics)] = pd.Series(
data=result
)
optimized_metrics.insert(loc=0, column="Name", value=["Old", "New"])
optimized_metrics = optimized_metrics.set_index("Name")
Оценка параметров старой и новой модели¶
In [71]:
optimized_metrics[
[
"Precision_train",
"Precision_test",
"Recall_train",
"Recall_test",
"Accuracy_train",
"Accuracy_test",
"F1_train",
"F1_test",
]
].style.background_gradient(
cmap="plasma",
low=0.3,
high=1,
subset=["Accuracy_train", "Accuracy_test", "F1_train", "F1_test"],
).background_gradient(
cmap="viridis",
low=1,
high=0.3,
subset=[
"Precision_train",
"Precision_test",
"Recall_train",
"Recall_test",
],
)
Out[71]:
In [72]:
optimized_metrics[
[
"Accuracy_test",
"F1_test",
"ROC_AUC_test",
"Cohen_kappa_test",
"MCC_test",
]
].style.background_gradient(
cmap="plasma",
low=0.3,
high=1,
subset=[
"ROC_AUC_test",
"MCC_test",
"Cohen_kappa_test",
],
).background_gradient(
cmap="viridis",
low=1,
high=0.3,
subset=[
"Accuracy_test",
"F1_test",
],
)
Out[72]:
In [74]:
_, ax = plt.subplots(1, 2, figsize=(10, 4), sharex=False, sharey=False
)
for index in range(0, len(optimized_metrics)):
c_matrix = optimized_metrics.iloc[index]["Confusion_matrix"]
disp = ConfusionMatrixDisplay(
confusion_matrix=c_matrix, display_labels=["Under 30", "30-40", "40-50", "50-60", "60-70", "70-80", "80+"]
).plot(ax=ax.flat[index])
plt.subplots_adjust(top=1, bottom=0, hspace=0.4, wspace=0.3)
plt.show()
Задача регрессии¶
In [3]:
from sklearn.model_selection import train_test_split
X = df.drop(columns=['Networth','Rank ', 'Name']) # Признаки
y = df['Networth'] # Целевая переменная для регрессии
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)
In [12]:
df
Out[12]:
In [ ]:
from sklearn.compose import ColumnTransformer
from sklearn.preprocessing import StandardScaler, OneHotEncoder
from sklearn.impute import SimpleImputer
from sklearn.pipeline import Pipeline
import pandas as pd
# Исправляем ColumnTransformer с сохранением имен колонок
columns_to_drop = []
num_columns = [
column
for column in X_train.columns
if column not in columns_to_drop and X_train[column].dtype != "object"
]
cat_columns = [
column
for column in X_train.columns
if column not in columns_to_drop and X_train[column].dtype == "object"
]
# Предобработка числовых данных
num_imputer = SimpleImputer(strategy="median")
num_scaler = StandardScaler()
preprocessing_num = Pipeline(
[
("imputer", num_imputer),
("scaler", num_scaler),
]
)
# Предобработка категориальных данных
cat_imputer = SimpleImputer(strategy="constant", fill_value="unknown")
cat_encoder = OneHotEncoder(handle_unknown="ignore", sparse_output=False, drop="first")
preprocessing_cat = Pipeline(
[
("imputer", cat_imputer),
("encoder", cat_encoder),
]
)
# Общая предобработка признаков
features_preprocessing = ColumnTransformer(
verbose_feature_names_out=True, # Сохраняем имена колонок
transformers=[
("prepocessing_num", preprocessing_num, num_columns),
("prepocessing_cat", preprocessing_cat, cat_columns),
],
remainder="drop" # Убираем неиспользуемые столбцы
)
# Итоговый конвейер
pipeline_end = Pipeline(
[
("features_preprocessing", features_preprocessing),
]
)
# Преобразуем данные
preprocessing_result = pipeline_end.fit_transform(X_train)
# Создаем DataFrame с правильными именами колонок
preprocessed_df = pd.DataFrame(
preprocessing_result,
columns=pipeline_end.get_feature_names_out(),
index=X_train.index, # Сохраняем индексы
)
preprocessed_df
Out[ ]:
In [13]:
import numpy as np
from sklearn.linear_model import LinearRegression
from sklearn.ensemble import RandomForestRegressor
from sklearn.ensemble import GradientBoostingRegressor
from sklearn.model_selection import GridSearchCV, RandomizedSearchCV
from sklearn.metrics import mean_absolute_error, mean_squared_error, r2_score
import matplotlib.pyplot as plt
random_state = 42
# Модели и параметры
models_regression = {
"LinearRegression": LinearRegression(),
"RandomForestRegressor": RandomForestRegressor(random_state=random_state),
"GradientBoostingRegressor": GradientBoostingRegressor(random_state=random_state)
}
param_grids_regression = {
"LinearRegression": {},
"RandomForestRegressor": {
'model__n_estimators': [10, 20, 30, 40, 50, 100, 150, 200, 250, 500],
'model__max_depth': [None, 2, 3, 4, 5, 6, 7, 8, 9 ,10],
},
"GradientBoostingRegressor": {
'model__n_estimators': [10, 20, 30, 40, 50, 100, 150, 200, 250, 500],
'model__learning_rate': [0.01, 0.1, 0.2, 0.3, 0.4, 0.5],
'model__max_depth': [2, 3, 4, 5, 6, 7, 8, 9 ,10]
}
}
# Результаты
results_regression = {}
# Перебор моделей
for name, model in models_regression.items():
print(f"Training {name}...")
pipeline = Pipeline(steps=[
('features_preprocessing', features_preprocessing),
('model', model)
])
param_grid = param_grids_regression[name]
grid_search = RandomizedSearchCV(pipeline, param_grid, cv=5, scoring='neg_mean_absolute_error', n_jobs=-1)
grid_search.fit(X_train, y_train)
# Лучшая модель
best_model = grid_search.best_estimator_
y_pred = best_model.predict(X_test)
# Метрики
mae = mean_absolute_error(y_test, y_pred)
rmse = np.sqrt(mean_squared_error(y_test, y_pred))
r2 = r2_score(y_test, y_pred)
# Сохранение результатов
results_regression[name] = {
"Best Params": grid_search.best_params_,
"MAE": mae,
"RMSE": rmse,
"R2": r2
}
# Печать результатов
for name, metrics in results_regression.items():
print(f"\nModel: {name}")
for metric, value in metrics.items():
print(f"{metric}: {value}")
In [14]:
# Импортируем pandas для работы с таблицами
import pandas as pd
# Формируем таблицу метрик
reg_metrics = pd.DataFrame.from_dict(results_regression, orient="index")[
["MAE", "RMSE", "R2"]
]
# Визуализация результатов с помощью стилизации
styled_metrics = (
reg_metrics.sort_values(by="RMSE")
.style.background_gradient(cmap="viridis", low=1, high=0.3, subset=["RMSE", "MAE"])
.background_gradient(cmap="plasma", low=0.3, high=1, subset=["R2"])
)
styled_metrics
Out[14]:
Шикарный вывод: по стране, возрасту, сфере деятельности и источнику доходов невозможно предсказать состояние человека. Значит ли это, что кто угодно, где угодно, и в чём угодно может добиться успеха?¶
Классификация¶
In [15]:
import pandas as pd
import matplotlib.pyplot as plt
df = pd.read_csv("C://Users//annal//aim//static//csv//Forbes_Billionaires.csv")
bins = [0, 30, 40, 50, 60, 70, 80, 101] # границы для возрастных категорий
labels = ['Under 30', '30-40', '40-50', '50-60', '60-70', '70-80', '80+'] # метки для категорий
df["Age_category"] = pd.cut(df['Age'], bins=bins, labels=labels, right=False)
# Удаляем оригинальные колонки 'country', 'industry' и 'source' из исходного DataFrame
df.drop(columns=['Age'], inplace=True)
# Просмотр результата
print(df.head())
In [18]:
from sklearn.ensemble import RandomForestClassifier
from sklearn.linear_model import LogisticRegression
from sklearn.neighbors import KNeighborsClassifier
from sklearn.metrics import accuracy_score, confusion_matrix, f1_score
X = df.drop(columns=['Age_category','Rank ', 'Name']) # Признаки
# Целевая переменная для классификации
y_class = df['Age_category']
# Разделение данных
X_train_clf, X_test_clf, y_train_clf, y_test_clf = train_test_split(X, y_class, test_size=0.2, random_state=42)
# Модели и параметры
models_classification = {
"LogisticRegression": LogisticRegression(max_iter=1000),
"RandomForestClassifier": RandomForestClassifier(random_state=42),
"KNN": KNeighborsClassifier()
}
param_grids_classification = {
"LogisticRegression": {
'model__C': [0.1, 1, 10]
},
"RandomForestClassifier": {
"model__n_estimators": [10, 20, 30, 40, 50, 100, 150, 200, 250, 500],
"model__max_features": ["sqrt", "log2", 2],
"model__max_depth": [2, 3, 4, 5, 6, 7, 8, 9 ,10, 20],
"model__criterion": ["gini", "entropy", "log_loss"],
},
"KNN": {
'model__n_neighbors': [3, 5, 7, 9, 11],
'model__weights': ['uniform', 'distance']
}
}
# Результаты
results_classification = {}
# Перебор моделей
for name, model in models_classification.items():
print(f"Training {name}...")
pipeline = Pipeline(steps=[
('features_preprocessing', features_preprocessing),
('model', model)
])
param_grid = param_grids_classification[name]
grid_search = RandomizedSearchCV(pipeline, param_grid, cv=5, scoring='f1', n_jobs=-1)
grid_search.fit(X_train_clf, y_train_clf)
# Лучшая модель
best_model = grid_search.best_estimator_
y_pred = best_model.predict(X_test_clf)
# Метрики
acc = accuracy_score(y_test_clf, y_pred)
f1 = f1_score(y_test_clf, y_pred)
# Вычисление матрицы ошибок
c_matrix = confusion_matrix(y_test_clf, y_pred)
# Сохранение результатов
results_classification[name] = {
"Best Params": grid_search.best_params_,
"Accuracy": acc,
"F1 Score": f1,
"Confusion_matrix": c_matrix
}
# Печать результатов
for name, metrics in results_classification.items():
print(f"\nModel: {name}")
for metric, value in metrics.items():
print(f"{metric}: {value}")