使用時間特征使讓機器學(xué)習(xí)模型更好地工作

來源：?DeepHub IMBA
本文約2300字，建議閱讀8分鐘
在本文中，通過一個實際示例討論如何從 DateTime 變量中提取新特征以提高機器學(xué)習(xí)模型的準(zhǔn)確性。

從日期中提取特征

• 一年中的一天或一個月中的一天或一周中的一天

• 一年中的月份

• 季節(jié)

• 年

import numpy as np
current_date = "2022-01-25 17:21:22"
cdate = datetime.strptime(current_date, '%Y-%m-%d %H:%M:%S')

day_sin = np.sin(2 * np.pi * cdate.timetuple().tm_yday/365.0)
day_cos = np.cos(2 * np.pi * cdate.timetuple().tm_yday/365.0)

df['month_sin'] = np.sin(2 * np.pi * df['date_time'].dt.month/12.0)
df['month_cos'] = np.cos(2 * np.pi * df['date_time'].dt.month/12.0)

df['year'] = df['date_time'].dt.year

從時間中提取特征

hour_sin = np.sin(2 * np.pi * cdate.hour/24.0)
hour_cos = np.cos(2 * np.pi * cdate.hour/24.0)

一個實際的例子

import pandas as pd
df = pd.read_csv('../input/weather-dataset/weatherHistory.csv')

df['Loud Cover'].value_counts()
0.0 ? 96453
Name: Loud Cover, dtype: int64

df.drop(['Daily Summary','Loud Cover'],axis=1,inplace=True)

df.isnull().sum()

Formatted Date ? ? ? ? ? ? ? 0
Summary ? ? ? ? ? ? ? ? ? ? ? 0
Precip Type ? ? ? ? ? ? ? ? 517
Temperature (C) ? ? ? ? ? ? ? 0
Apparent Temperature (C) ? ? 0
Humidity ? ? ? ? ? ? ? ? ? ? 0
Wind Speed (km/h) ? ? ? ? ? ? 0
Wind Bearing (degrees) ? ? ? 0
Visibility (km) ? ? ? ? ? ? ? 0
Pressure (millibars) ? ? ? ? 0

df.dropna(inplace=True)

from sklearn.preprocessing import LabelEncoder

le = LabelEncoder()
df['Precip Type']=le.fit_transform(df['Precip Type'])
df['Summary']=le.fit_transform(df['Summary'])

from sklearn.preprocessing import StandardScaler
scaler = StandardScaler()
df[df.columns[2:]] = scaler.fit_transform(df[df.columns[2:]])

import numpy as np
from datetime import datetime

def discretize_date(current_date, t):
? current_date = current_date[:-10]
? cdate = datetime.strptime(current_date, '%Y-%m-%d %H:%M:%S')
? ?
? if t == 'hour_sin':
? ? ? return np.sin(2 * np.pi * cdate.hour/24.0)
? if t == 'hour_cos':
? ? ? return np.cos(2 * np.pi * cdate.hour/24.0)
? if t == 'day_sin':
? ? ? return np.sin(2 * np.pi * cdate.timetuple().tm_yday/365.0)
? if t == 'day_cos':
? ? ? return np.cos(2 * np.pi * cdate.timetuple().tm_yday/365.0)

date_types = ['hour_sin', 'hour_cos', 'day_sin', 'day_cos']
for dt in date_types:
? df[dt] = df['Formatted Date'].apply(lambda x : discretize_date(x, dt))
df.drop(['Formatted Date'],axis=1,inplace=True)

df.corr()

df.drop(['Apparent Temperature (C)'],axis=1,inplace=True)

from sklearn.model_selection import train_test_split
X = df.iloc[:,1:]
y=df.iloc[:,0]?

X_train,X_test,y_train,y_test = train_test_split(X,y,test_size=0.3,random_state=1)

from sklearn.ensemble import RandomForestClassifier

model = RandomForestClassifier(max_depth=32,n_estimators=120,random_state=1)
model.fit(X_train,y_train)
y_pred = model.predict(X_test)

from sklearn.metrics import accuracy_score
accuracy_score(y_test, y_pred)
0.6695389319342622

model = RandomForestClassifier(max_depth=32,n_estimators=120,random_state=1)
model.fit(X_train[X_train.columns[:-4]],y_train)
y_pred = model.predict(X_test[X_test.columns[:-4]])
accuracy_score(y_test, y_pred)
0.5827108161634411

總結(jié)