python數據幀的時間序列預測

Question

我正在處理如下所示的代碼：

df=pd.read_csv("file.csv")
df['fraction'] = df ['number'] / df['year_total']
df.fraction = df.fraction.round(4)
df

給出輸出為

programming_lang = ["r", "python", "c#", "java", "JavaScript", "php", "c++", "ruby", "Selenium"]

yearly_top = df[df['tag'].isin(programming_lang)]
yearly_top

給出如下輸出：

year, tag, number, year_total, fraction
2008, java, 7473, 58390,0.1280
2008, php, 3111, 58390, 0.0533
2008, Python, 2080, 58390, 0.0356
......
2019, java, 83841, 1085170, 0.0773
2019, php, 61257, 1085170, 0.0564
2019, python, 107348, 1085170, 0.0989

它包含了 2008 年到 2019 年的頂級編程語言數據。我想使用時間序列模型來預測這些編程語言在 2020 年、2021 年和 2022 年的fraction值。我對這個領域很陌生。 任何線索都會有所幫助

Answer 1

您可以使用 RNN 解決它。 首先，讓我們創建一個示例數據框來使用

import pandas as pd 
import numpy as np

test_df = pd.DataFrame({'year':range(2008,2020)})
# 0-java, 1-php, 2-python
for ind in range(3): test_df['frac_%i' % ind] = np.random.rand(2020-2008)
test_df = test_df.drop('year',axis=1)
# the array of fractions 
data = test_df.values

在刪除列year之前， test_df看起來像

    year    frac_0    frac_1    frac_2
0   2008  0.457123  0.780754  0.978396
1   2009  0.578795  0.323664  0.909824
2   2010  0.707996  0.477242  0.948976
3   2011  0.455918  0.627572  0.137039
4   2012  0.272352  0.144968  0.831693
5   2013  0.064729  0.233168  0.554654
6   2014  0.754608  0.570530  0.968355
7   2015  0.435918  0.264335  0.727189
8   2016  0.699624  0.455323  0.237246
9   2017  0.824758  0.995260  0.333113
10  2018  0.597993  0.384319  0.750074
11  2019  0.598657  0.533934  0.072334

在使用 RNN 進行時間序列分析時，首先要將任務轉換為監督回歸任務，即我們需要創建一個數據框，其中每一行都是

observations of the past year | observation of a year

這是一個可以幫助您實現此目的的功能（我從這篇精彩的帖子中學到了這個功能）

def series_to_supervised(data,n_in,n_out):

    df = pd.DataFrame(data)
    cols = list()
    for i in range(n_in,0,-1): cols.append(df.shift(i))
    for i in range(0, n_out): cols.append(df.shift(-i))
    agg = pd.concat(cols,axis=1)
    agg.dropna(inplace=True)

    return agg.values

使用此函數，我們可以創建所需的數據幀

n_in,n_out = 2,1
data = series_to_supervised(test_df,n_in,n_out)

n_in是我們想要用來進行預測的過n_out數， n_out是我們想要預測的年數。 在這種情況下，我們僅根據過去兩年的數據預測一年。

現在我們已經准備好了數據，我們可以訓練一個 RNN 模型

from sklearn.model_selection import train_test_split
from tensorflow.keras.models import Model,Sequential
from tensorflow.keras.layers import Dense,LSTM,Dropout

x, y= data[:,None,:-n_out*3],data[:,n_in*3:]
x_train,x_test,y_train,y_test = train_test_split(x,y,test_size=0.1,random_state=49)

model = Sequential()
model.add(LSTM(4,name='lstm_0'))
model.add(Dropout(0.2,name='dropout_0'))
model.add(Dense(3,activation='tanh'))
model.compile(loss='mse',optimizer='adam',metrics=['mse'])
# fit
history = model.fit(x_train,y_train,validation_data=(x_test,y_test),epochs=50,verbose=0)

使用此模型，您可以通過以下方式預測 2020、2021 和 2022 年的分數

# predict 2020 with 2018 and 2019
last_two_years = np.hstack((test_df.values[-2],test_df.values[-1]))[None,None,:]
frac_2020 = model.predict(last_two_years)
# predict 2021 with 2019 and 2020
last_two_years = np.hstack((test_df.values[-1],frac_2020.ravel()))[None,None,:]
frac_2021 = model.predict(last_two_years)
# predict 2022 with 2020 and 2021
last_two_years = np.hstack((frac_2020.ravel(),frac_2021.ravel()))[None,None,:]
frac_2022 = model.predict(last_two_years)

完整的腳本

import pandas as pd 
import numpy as np
from sklearn.model_selection import train_test_split
from tensorflow.keras.models import Model,Sequential
from tensorflow.keras.layers import Dense,LSTM,Dropout


def series_to_supervised(data,n_in,n_out):

    df = pd.DataFrame(data)
    cols = list()
    for i in range(n_in,0,-1): cols.append(df.shift(i))
    for i in range(0, n_out): cols.append(df.shift(-i))
    agg = pd.concat(cols,axis=1)
    agg.dropna(inplace=True)

    return agg.values

test_df = pd.DataFrame({'year':range(2008,2020)})
# 0-java, 1-php, 2-python
for ind in range(3): test_df['frac_%i' % ind] = np.random.rand(2020-2008)
test_df = test_df.drop('year',axis=1)
# the array of fractions 
data = test_df.values
# cast the task as a supevised regression task 
n_in,n_out = 2,1
data = series_to_supervised(test_df,n_in,n_out)
# train test split
x, y= data[:,None,:-n_out*3],data[:,n_in*3:]
x_train,x_test,y_train,y_test = train_test_split(x,y,test_size=0.1,random_state=49)

model = Sequential()
model.add(LSTM(4,name='lstm_0'))
model.add(Dropout(0.2,name='dropout_0'))
model.add(Dense(3,activation='tanh'))
model.compile(loss='mse',optimizer='adam',metrics=['mse'])
# fit
history = model.fit(x_train,y_train,validation_data=(x_test,y_test),epochs=50,verbose=0)

# predict 2020 with 2018 and 2019
last_two_years = np.hstack((test_df.values[-2],test_df.values[-1]))[None,None,:]
frac_2020 = model.predict(last_two_years)
# predict 2021 with 2019 and 2020
last_two_years = np.hstack((test_df.values[-1],frac_2020.ravel()))[None,None,:]
frac_2021 = model.predict(last_two_years)
# predict 2022 with 2020 and 2021
last_two_years = np.hstack((frac_2020.ravel(),frac_2021.ravel()))[None,None,:]
frac_2022 = model.predict(last_two_years)

print(frac_2020,frac_2021,frac_2022)