WeightedCorr

Weighted correlation in Python. Pandas based implementation of weighted Pearson and Spearman correlations.

v2.1 20-03-2021

Fixed Issue #1

V2 Update 21-07-2020

Switched from a pandas backend to a numpy/scipy backend. Usage remains the same, but performance for Spearman correlations is significantly improved. See table below.

Install

pip install wcorr

Usage

This class can be used in a few different ways depending on your needs. The data should be passed to the initialization of the class. Then calling the class will produce the result with desired method (pearson is the default). Note that the method should be passed to the call, not the initialization. The examples below will result in pearson, pearson, and spearman correlations.

from wcorr import WeightedCorr

1. You can supply a pandas DataFrame with x, y, and w columns (columns should be in that order). The output will be a single floating point value.

WeightedCorr(xyw=my_data[['x', 'y', 'w']])(method='pearson')

2. You can supply x, y, and w pandas Series separately. The output will be a single floating point value.

WeightedCorr(x=my_data['x'], y=my_data['y'], w=my_data['w'])()

3. You can supply a pandas DataFrame, and the name of the weight column in that DataFrame. In this case the output will be an (M-1)x(M-1) pandas DataFrame (the correlation matrix) where M is the number of columns in the original dataframe (no correlation is calculated for the weight column, hence M-1).

WeightedCorr(df=my_data, wcol='w')(method='pearson')

Weighted Pearson correlation

The weighted Pearson r, given n pairs is calculated as

Where

Weighted Spearman rank-order correlation

First, initial ranks (z) are assigned to x and y. Duplicate groups of records are assigned the average rank of that group. Next the weighted rank (rank) is calculated for x and y separately in n pairs. Such that the j-th rank of either x or y will be:

Where

and

These weighted ranks are then passed to the weighted Pearson correlation function.