如何使用 python pandas 打印相关特征？

Question

I'm trying to get some information on the correlation of the independent variables.

My dataset has a lot of variables, therefore the heatmap is not solution, it is very unreadable.

Currently, I have made a function that returns only those variables that are highly correlated. I would like to change it in way to indicate pairs of correlated features.

The rest of the explanations below:

def find_correlated_features(df, threshold, target_variable):

    df_1 = df.drop(target_variable)

    #corr_matrix has in index and columns names of variables
    corr_matrix = df_1.corr().abs()

    # I'm taking only half of this matrix to prevent doubling results
    half_of_matrix = corr_matrix.where(np.triu(np.ones(corr_matrix.shape), k = 1).astype(np.bool))

    # This prints list of columns which are correlated 
    to_drop = [column for column in half_of_matrix.columns if any(half_of_matrix[column] > threshold)]
    
    return to_drop 

The best if this function would return pandas dataframe with column_1; column_2; corr_coef only variables that are above threshold.

Something like this:

output = {'feature name 1': column_name,
          'feature name 2': index,
          'correlation coef': corr_coef}

output_list.append(output)
return pd.DataFrame(output_list).sort_values('corr_coef', ascending=False)

Answer 1

This should match the output you're looking for:

import pandas as pd
import numpy as np

# Create fake correlation matrix
corr_matrix = np.random.random_sample((5, 5))


ii, jj = np.triu_indices(corr_matrix.shape[0], 1)

scores = []

for i, j in zip(ii, jj):
    scores.append((i,j,corr_matrix[i,j]))

df_out = pd.DataFrame(data=scores,columns=['feature name 1','feature name 2','corr_coef'])\
           .sort_values('corr_coef', ascending=False)\
           .reset_index(drop=True)

threshold = 0.1

df_out[df_out['corr_coef'] > threshold]

#  feature name 1 feature name 2    corr_coef
# 0     0              2            0.990691
# 1     2              4            0.990444
# 2     0              1            0.830640
# 3     1              2            0.623895
# 4     1              4            0.433258
# 5     3              4            0.404395
# 6     0              4            0.291564
# 7     2              3            0.276799
# 8     1              3            0.177519

And you can map the indices of the features (in columns feature name 1 and feature name 2 above) to the columns of your df_1 to get the actual feature names.

So your complete function would look like this:

def find_correlated_features(df, threshold, target_variable):

    df_1 = df.drop(target_variable)

    #corr_matrix has in index and columns names of variables
    corr_matrix = df_1.corr().abs().to_numpy()

    ii, jj = np.triu_indices(corr_matrix.shape[0], 1)

    scores = []

    for i, j in zip(ii, jj):
        scores.append((i,j,corr_matrix[i,j]))

    df_out = pd.DataFrame(data=scores,columns=['feature name 1','feature name 2','corr_coef'])\
               .sort_values('corr_coef', ascending=False)\
               .reset_index(drop=True)

    # This should go from the second column as the index loop that gave us
    # the scores and indices were from the upper triangle offset by 1
    feature_name_map = {i:c for i,c in enumerate(df_1.columns[1:])}

    df_out['feature name 1'] = df_out['feature name 1'].map(feature_name_map)
    df_out['feature name 2'] = df_out['feature name 2'].map(feature_name_map)

    return df_out[df_out['corr_coef'] > threshold] 

Answer 2

After Edit:

After OP comment and @user6386471 answer, I've read again the question and I think that a simply restructure of the correlation matrix would work, with no need of loops. Like half_of_matrix.stack().reset_index() plus filters. See:

def find_correlated_features(df, threshold, target_variable):
    # remove target column
    df = df.drop(columns=target_variable).copy()
    # Get correlation matrix
    corr_matrix = df.corr().abs()
    # Take half of the matrix to prevent doubling results
    corr_matrix = corr_matrix.where(np.triu(np.ones(corr_matrix.shape), k = 1).astype(np.bool))
    # Restructure correlation matrix to dataframe
    df = corr_matrix.stack().reset_index()
    df.columns = ['feature1', 'feature2', 'corr_coef']
    # Apply filter and sort coefficients
    df = df[df.corr_coef >= threshold].sort_values('corr_coef', ascending=False)
    return df

---

Original answer:

You can easily create a Series with the coefficients above a threshold like this:

s = df.corr().loc[target_col]
s[s.abs() >= threshold]

where df is your dataframe, target_col your target column, and threshold , you know, the threshold.

---

Example:

import pandas as pd
import seaborn as sns

df = sns.load_dataset('iris')

print(df.shape)
# -> (150, 5)

print(df.head())

   sepal_length  sepal_width  petal_length  petal_width species
0           5.1          3.5           1.4          0.2  setosa
1           4.9          3.0           1.4          0.2  setosa
2           4.7          3.2           1.3          0.2  setosa
3           4.6          3.1           1.5          0.2  setosa
4           5.0          3.6           1.4          0.2  setosa

def find_correlated_features(df, threshold, target_variable):
    s = df.corr().loc[target_variable].drop(target_variable)
    return s[s.abs() >= threshold]

find_correlated_features(df, .7, 'sepal_length')

output:

petal_length    0.871754
petal_width     0.817941
Name: sepal_length, dtype: float64

You can use .to_frame() followed by .T to the outptut to get a pandas dataframe.