Basics, Python

Introduction to Linear Discriminant Analysis

Linear Discriminant Analysis

Linear Discriminant Analysis or most commonly known as ‘LDA’ is one of the most interesting machine learning techniques till date.The idea was first coined by “Dr. Ronald Fisher” to classify binary classes using ‘Fisher’s linear discriminant‘ and later on it was generalized for multiple classes as well.In case of a binary class problem, LDA acts as a classifier like “logistic regression” and when it comes to multi classes it acts as a dimensional reduction technique like “PCA”.In this article I will not be talking much about how to use LDA for a multi class problem, instead, let’s focus on binary classes. However, let’s take a peak into how “LDA” and “PCA” are different yet function the same.

LDA- linear discriminant Analysis

Looking at the figure above lets try to draw some conclusions.It seems that both PCA and LDA are trying to project the data points in a vector space(in this case a straight line), thus reducing the dimension from ‘2’ to ‘1’. However, in case of PCA the points are projected along a vector space(line) whose direction is aligned to where scatteredness is maximum, but for LDA its the opposite.There is one more difference between the two, in LDA for a multiclass problem the number of LDA components  will  always be ‘C-1′ where ‘C’ represents the number of classes, unlike PCA  where it is independent of the classes. For the time being lets forget about PCA and concentrate on LDA. There are a lot of differences between the two but lets not go into that and concentrate on the task in our hand i.e use LDA to classify a binary class problem.We will not dive into the mathematical details of ‘LDA’ but I will give you certain formulas which will help you to apply the LDA technique on any binary classification problem.

Using Linear Discriminant Analysis on Binary Classification 

Step 1: Create a matrix containing the mean of each of the features (X) pertaining to a particular class say ‘Ci’.


Step 2: Create a scatter matrix (Si) for each class using the following formula.

eqn 2

Step 3: Calculate the projection vector(W).

Eqn. 3

Step 4: The last and final step is to project these data points in this vector (W).Here ‘Y’ represents the new projected data points which is a scalar quantity.

Eqn. 4

Let’s look at the following figure of a two feature binary classification problem.You can clearly see how the data points are projected along the vector(the line represents W) after applying LDA.

Introduction to LDA (Linear Discriminant Analysis)

I have given you the basic difference between “LDA” and “PCA“, also I have given all the necessary steps for LDA but the question remains, How is LDA doing what it’s doing?  I haven’t given you the intuition which led ‘Dr.Fisher‘ to formulate this beautiful technique. If you recall “logistic regression” what it does is that it takes a number of input features and represents in terms of probabilities.Well, LDA does the same except it takes the input feature vector(don’t get confused it means the set of features) and represents them in terms of scalar quantity.Yes, you read that right it converts vectors of features to scalars.Now, these scalars can be used to separate the two classes and very soon we will go through an example to get a better idea of the same.But first, let’s look at the following figure.

Fisher's linear discriminant

I hope the above figure cleared your doubts a little, but let me try to explain in a bit detail.The function ‘J(W)’ is the one we want to maximize, but what is it maximizing.From the function, it looks like the numerator i.e the difference between the means of the two classes should be maximized while keeping the scatteredness as minimum as possible.So the direction of the vector ‘W’ should be such that the all the above criteria are satisfied.Our objective is to find this vector ‘W’ first as shown in step 3 and finally use it to get the scalar quantities for each set of input feature vector(step 4).Let’s code the above steps in Python and see how it turns out.


Linear Discriminant Analysis for Binary Classification

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from pandas import Series,DataFrame
%matplotlib inline

First of all lets prepare a data set

data = DataFrame()
X1 = Series(np.array([4,2,2,3,4,9,6,9,8,10]))
X2 = Series(np.array([1,4,3,6,4,10,8,5,7,8]))
 Lets take a look at the data

Lets plot the above data and see how it looks

<matplotlib.legend.Legend at 0x2090c46a908>

Using Step 1 to calculate the mean matrix

mu1 = np.array([np.mean(data.ix[data['class']=='class1',0]),np.mean(data.ix[data['class']=='class1',1])])
mu2 = np.array([np.mean(data.ix[data['class']=='class2',0]),np.mean(data.ix[data['class']=='class2',1])])
[ 3.   3.6] [ 8.4  7.6]

Using Step 2 to find the scatter mattrices of each class and also the within class scatter matrix

s1 =[data['class']=='class1',:-1])-mu1).T,np.array(data.ix[data['class']=='class1',:-1])-mu1)
s1#scatter matrix for class 1
array([[  4. ,  -2. ],
       [ -2. ,  13.2]])
s2 =[data['class']=='class2',:-1])-mu2).T,np.array(data.ix[data['class']=='class2',:-1])-mu2)
s2#scatter matrix for class 2
array([[  9.2,  -0.2],
       [ -0.2,  13.2]])
sw = s1 + s2
sw#within class scatter matrix
array([[ 13.2,  -2.2],
       [ -2.2,  26.4]])

Using Step 3 to find the projection vector W

W =,(mu1-mu2).reshape(2,1))
W = -W.T[0] #converting it to positive as it is not going to create much of a difference
array([ 0.44046095,  0.18822023])

Using Step 4 to find the scalar quantities for each set of input features

for i in range(len(data.ix[:,:-1])):

The final data along with the projected scalar quantities looks like this


Lets plot and see how well they are separated

plt.scatter(data.ix[data['class']=='class1',3],np.array([1,1,1,1,1]),color='r',marker = '*',label='class1')
plt.scatter(data.ix[data['class']=='class2',3],np.array([1,1,1,1,1]),color='g',marker = 'o',label='class2')
<matplotlib.legend.Legend at 0x2090caa7588>


Using the above plot of the projected scalar quantities we can easily separate the two classes i.e any class with a scalar value less than, say ‘3’ belongs to class 1 and values greater than ‘4’ belongs to class 2.How easily LDA solved our problem of binary classification, but I wish life was that simple.There might be some cases that even after applying LDA there is overlapping of the two classes.What to do then? Our model’s accuracy will degrade.Is there a solution? Well, there is, we will have to combine LDA with some other techniques which I will show you in some other article, for time being let’s hope that our classes are always well separated.Hope you enjoyed the article!!!