Home » Machine Learning/Artificial Intelligence

# Kendall's Tau correlation in machine learning

In this article, we would be looking at **Kendall’s correlation and why should we use this for machine learning**?

Submitted by Raunak Goswami, on October 05, 2018

Before we begin I hope you guys have a basic understanding of Pearson’s and Spearman's correlation. As the name suggests this correlation was named after **Maurice Kendall in the year 1938**.

This type of correlation is best suited for the discrete data. Here we are not completely dependent on the directional flow of the ranks of various observation that we used to do in spearman’s correlation. Here we are more concerned with concordant pairs and discordant pairs.

**1. Concordant pairs**

For a given set of data the concordant pairs are such that for a given set of data suppose **(x1, y1) and (x2, y2)** then **x1<x2** and **y1<y2** where **x1** and **x2** can be any of the attribute values and **y1** and **y2** are the values in the target column.

**2. Discordant pairs**

For a given set of data, the discordant pairs would be the pairs which do not satisfy the property of the concordant pairs which is **x1<x2** and **y1<y2**. Where **x1** and **x2** can be any of the attribute values and **y1** and **y2** are the values in the target column.

After calculating concordant and discordant pairs we find the difference between them and then divide the result by the number of possible combinations of the different pairs. The main aim of dividing the difference by the number of possible combination pairs is to make the value of Kendall's coefficient i.e. tau to fall under -1 to 1 so that it is easier to find out whether the given attribute should be used for predictive analysis of the target value. Unlike other correlations here too, 0 will signify 0 correlation and 1 signifies perfect correlation and -1 signifies the negative correlation.

The mathematical formula for the given correlation is mentioned below:

((Number of concordant pairs) - (number of discordant pairs))/(N(N-1))/2

Here, **(N(N-1))/2** is the number of possible pairs in the dataset

**Dataset description:**

The data set used has two columns i.e.

- YearsExperience
- Salary

The data set tell about the salary of the different employees based on the years of experience in their field so we would be using correlation to find out the relation between years of experience and the salary.

The data set can be downloaded from here: **Salary_Data.csv**

Now without wasting any time let us write the python code for the following correlation.

**Code:**

# -*- coding: utf-8 -*- """ Created on Sun Jul 29 22:21:12 2018 @author: Raunak Goswami """ import numpy as np import pandas as pd import matplotlib.pyplot as plt #reading the data """ here the directory of my code and the headbrain4.csv file is same make sure both the files are stored in the same folder or directory """ data=pd.read_csv('Salary_Data.csv') #this will show the first five records of the whole data data.head() #this will create a variable w which has the feature values i.e years of experience w=data.iloc[:,0:1].values #this will create a variable x which has the feature values i.e salary y=data.iloc[:,1:2].values print(round(data['YearsExperience'].corr(data['Salary'],method='kendall'))) plt.scatter(w,y,c='red') plt.title('scattered graph for kendall correlation between years of experience and salary' ) plt.xlabel('Gender') plt.ylabel('brain weight') plt.show() data.info() data['YearsExperience'].corr(data['Salary']) k1=data.corr(method='kendall') print("The table for all possible values of kendall'scoeffecients is as follows") print(k1)

**Output**

From the given output the value of **Kendall tau’s correlation coefficient between years of experience and salary comes out to be 0.841016** which is a fairly good correlation value. That was all for today guys hope you liked this article. Keep learning.

Comments and Discussions

**Ad:**
Are you a blogger? Join our Blogging forum.