CODEWITHSUNDEEP
pythonpandasnumpymatrixvectorization
Clade
Clade

Reputation: 986

Create a numpy covariance matrix from a pandas DataFrame of covariances

I have the following pandas.DataFrame object that provides the covariances between factors:

import pandas as pd

df = pd.DataFrame({"factor1": ["A", "A", "A", "B", "B", "C"],
                   "factor2": ["A", "B", "C", "B", "C", "C"],
                   "covar": [-1.2, -1, 2, 3.4, -4, 6.2]})

My objective is to reformat the DataFrame into a positive semi-definite covariance numpy.ndarray.

I have developed a working solution, however, it is painfully slow:

unique_factors = df.factor1.unique()
F = pd.DataFrame(columns=unique_factors, index=unique_factors)
for index, row in df.iterrows():
    F.loc[row["factor1"], row["factor2"]] = row["covar"]**2
    F.loc[row["factor2"], row["factor1"]] = row["covar"]**2 #inefficient
F = F.to_numpy()

The output of which is:

[[1.44 1.0                4.0               ]
 [1.0  11.559999999999999 16.0              ]
 [4.0  16.0               38.440000000000005]]

I am hoping that I can take advantage of numpy's native methods to accomplish my objective more efficiently. At the very least I would like to be able to remove the line commented #inefficient and reflect the upper triangular matrix about the diagonal. Any help would be much appreciated.

Upvotes: 0

Views: 986

Answers (1)

BENY
BENY

Reputation: 323226

In your case 

s=df.pivot(*df.columns)**2
s=s.fillna(s.T)

Out[230]: 
factor2     A      B      C
factor1                    
A        1.44   1.00   4.00
B        1.00  11.56  16.00
C        4.00  16.00  38.44

Upvotes: 2

Related Questions