[FIXED] Is there a way to extract the important features from XGBoost automatically and use for prediction?

Issue

I am trying to develop a prediction model using XGBoost. My basic idea is to develop an automated prediction model which uses the top 10 important features derived from the dataset (700+ rows and 90+columns) and use them for prediction of values.

The input data is updated weekly and hence the predictions for the next week should be predicted using current week values. I have extracted important features from my XGBoost model but am unable to automate the same due to the error.

import xgboost as xgb
from sklearn.metrics import mean_squared_error
from sklearn.metrics import accuracy_score
from sklearn.model_selection import train_test_split

X_train, X_test, y_train, y_test = train_test_split(X, Y, test_size=0.2, random_state=100)
eval_set = [(X_train, y_train), (X_test, y_test)]

xg_reg = MyXGBRegressor(objective ='reg:squarederror', colsample_bytree = 0.3, learning_rate = 0.01,max_depth = 6, reg_alpha = 15, n_estimators = 1000, subsample = 0.5)
predictions = xg_reg.fit(X_train,y_train, early_stopping_rounds=30, eval_metric=["rmse", "mae"], eval_set=eval_set, verbose=True)

The above code helps me run the regressor and predict values. The following code throws an error.

import xgboost as xgb
from xgboost import XGBRegressor

class MyXGBRegressor(XGBRegressor):
    @property
    def coef_(self):

    return None

thresholds = np.sort(xg_reg.feature_importances_)

from sklearn.feature_selection import SelectFromModel

for thresh in thresholds:
    selection = SelectFromModel(xg_reg, threshold=thresh, prefit = True)
    selected_dataset = selection.transform(X_test)
    feature_idx = selection.get_support()
    feature_name = X.columns[feature_idx]
    selected_dataset = pd.DataFrame(selected_dataset)
    selected_dataset.columns = feature_name

The error is as follows:

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
<ipython-input-11-a42c3ed80da2> in <module>
      3 for thresh in thresholds:
      4     selection = SelectFromModel(xg_reg, threshold=thresh, prefit = True)
----> 5     selected_dataset = selection.transform(X_test)
      6 
      7 feature_idx = selection.get_support()

~\Anaconda3\lib\site-packages\sklearn\feature_selection\_base.py in transform(self, X)
     86             force_all_finite=not _safe_tags(self, key="allow_nan"),
     87         )
---> 88         mask = self.get_support()
     89         if not mask.any():
     90             warn("No features were selected: either the data is"

~\Anaconda3\lib\site-packages\sklearn\feature_selection\_base.py in get_support(self, indices)
     50             values are indices into the input feature vector.
     51         """
---> 52         mask = self._get_support_mask()
     53         return mask if not indices else np.where(mask)[0]
     54 

~\Anaconda3\lib\site-packages\sklearn\feature_selection\_from_model.py in _get_support_mask(self)
    186                              ' "prefit=True" while passing the fitted'
    187                              ' estimator to the constructor.')
--> 188         scores = _get_feature_importances(
    189             estimator=estimator, getter=self.importance_getter,
    190             transform_func='norm', norm_order=self.norm_order)

~\Anaconda3\lib\site-packages\sklearn\feature_selection\_base.py in _get_feature_importances(estimator, getter, transform_func, norm_order)
    189         return importances
    190     elif transform_func == "norm":
--> 191         if importances.ndim == 1:
    192             importances = np.abs(importances)
    193         else:

AttributeError: 'NoneType' object has no attribute 'ndim'

Solution

The problem is that the coef_ attribute of MyXGBRegressor is set to None. If you use XGBRegressor instead of MyXGBRegressor then SelectFromModel will use the feature_importances_ attribute of XGBRegressor and your code will work.

import numpy as np
from xgboost import XGBRegressor
from sklearn.datasets import make_regression
from sklearn.model_selection import train_test_split
from sklearn.feature_selection import SelectFromModel

# generate some data
X, y = make_regression(n_samples=1000, n_features=5, random_state=100)

# split the data
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=100)

# instantiate the model
model = XGBRegressor(objective="reg:squarederror", colsample_bytree=0.3, learning_rate=0.01, max_depth=6, reg_alpha=15, n_estimators=1000, subsample=0.5)

# fit the model
model.fit(X_train, y_train, early_stopping_rounds=30, eval_metric=["rmse", "mae"], eval_set=[(X_train, y_train), (X_test, y_test)], verbose=True)

# extract the feature importances
thresholds = np.sort(model.feature_importances_)

# select the features
selection = SelectFromModel(model, threshold=thresholds[2], prefit=True)

feature_idx = selection.get_support()
print(feature_idx)
# array([ True,  True,  True, False, False])

selected_dataset = selection.transform(X_test)
print(selected_dataset.shape)
# (200, 3)

Answered By - Flavia Giammarino