About get_coef in older version

[kennis222]

Hi developers,

The function "get_coef" is deprecated since version 4.3.0, but I would like to use the get_coef with an older version since it easier to interpret to people without machine learning or statistical knowledge compared to use the "impurity-based feature importance" . However, when I use the function "get_coef" in the older version either 4.2.0 or 4.1.0, the program reports error with "AttributeError: 'ForecasterAutoregMultiOutput' object has no attribute 'get_coef'"

Here is the codes I calls for the function:
coef_ = []
for i,j in zip(range(1,7),pd.date_range(start='2022-01',periods=6,freq='M')):
<space>df = pd.DataFrame()
<space>df['feature'] = model_direct.get_coef(step=i)['feature']
<space>df['coef'] = model_direct.get_coef(step=i)['coef']
<space> df['date'] = pd.to_datetime(j)
<space>coef_.append(df)

Thanks.

[JavierEscobarOrtiz]

Hello @kennis222,

Thank you for opening the issue, a similar example works for me in version 0.4.3, here is:

# Libraries
# ==============================================================================
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from skforecast.ForecasterAutoregMultiOutput import ForecasterAutoregMultiOutput
from sklearn.linear_model import LinearRegression
from sklearn.metrics import mean_squared_error

# Create and fit forecaster
# ==============================================================================
y = pd.Series(list(range(50)))

forecaster = ForecasterAutoregMultiOutput(
                 regressor = LinearRegression(),
                 steps     = 5,
                 lags      = 2
             )

forecaster.fit(y=y)

# Get coef for step 1
# ==============================================================================
display(forecaster.get_coef(step=1))

# Get list of df
# ==============================================================================
coef_ = []

for i in list(range(forecaster.steps)):
    df = forecaster.get_coef(step=i+1) # First step is 1 not 0
    coef_.append(df)
    
# print Get coef for step 1
display(coef_[0])

From forecaster.get_coef(step=1):

	feature	coef
0	lag_1	0.5
1	lag_2	0.5

From list:

	feature	coef
0	lag_1	0.5
1	lag_2	0.5

FYI, the get_coef method was deprecated because it was merged into the get_feature_importance method. get_feature_importance accesses the method coef_ or feature_importances_ depending on the type of the regressor.

So in the latest version, 0.5.1 it would be:

Note that ForecasterAutoregMultiOutput has been renamed to ForecasterAutoregDirect

# Libraries
# ==============================================================================
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from skforecast.ForecasterAutoregDirect import ForecasterAutoregDirect
from sklearn.linear_model import LinearRegression
from sklearn.metrics import mean_squared_error

# Create and fit forecaster
# ==============================================================================
y = pd.Series(list(range(50)))

forecaster = ForecasterAutoregDirect(
                 regressor = LinearRegression(),
                 steps     = 5,
                 lags      = 2
             )

forecaster.fit(y=y)

# Get coef for step 1
# ==============================================================================
display(forecaster.get_feature_importance(step=1))

# Get list of df
# ==============================================================================
coef_ = []

for i in list(range(forecaster.steps)):
    df = forecaster.get_feature_importance(step=i+1) # First step is 1 not 0
    coef_.append(df)
    
# print Get coef for step 1
display(coef_[0])

From forecaster.get_feature_importance(step=1):

	feature	importance
0	lag_1	0.5
1	lag_2	0.5

From list:

	feature	importance
0	lag_1	0.5
1	lag_2	0.5

I recommend using the latest version as we have fixed some bugs and improved efficiency. 😄

Thank you!

[kennis222]

Hello @JavierEscobarOrtiz,

Thank you for the reply, but the case you show me in the version 0.4.3 still not works for my environment. However, since merging the the method coef_ or feature_importances_ depending on the type of the regressor, can I understand in this way: if using LinearRegression, the feature_importances_ means the coefficients, but if using the random forest, the feature_importances_ means the impurity-based feature importance? If "Yes", I can directly use the latest version : )

[JoaquinAmatRodrigo]

Hi @kennis222,
Yes, you understood it well :)