Implementing logistic regression learner with python

Logistic regression is a next step from linear regression. The most real life data have non linear relationship, thus applying linear models might be ineffective. Logistic regression is capable of handling hon linear effects in prediction tasks. You can think of lots of different scenarios where logistic regression could be applied. There can be financial, demographic, health, weather and other data where model could be applied and used to predict next events on upcoming data. For instance you can classify emails in to span and non spam, transactions being fraud or non, tumors being malignant or benign.

In order to understand logistic regression, let’s cover some basics, do a simple classification on data set with two features and then test it on real life data with multiple features.

Sigmoid function kernel

Since logistic regression based classifier is non linear, we need a non linear kernel function. Logistic regression uses sigmoid function which is “S” shaped curve. It can have values from 0 to 1 which is convenient when deciding to which class assign the output value.

Using python we can draw sigmoid graph:

import numpy as np
import matplotlib.pyplot as plt
z = np.arange(-6, 6, 0.1);
sigmoid = 1/(1+np.exp(-z));
fig = plt.figure('Cost function convergence')
plt.xlabel(' input z')
plt.title('Sigmoid function graph')

To make a hypothesis out of sigmoid we need to substitute z parameter with vectorized form of ΘTx which represents input features and parameters (weights). Where,

ΘTx = Θ01x12x2+…+Θnxn

Note: ΘTx can be constructed differently using mapping of features to meet training set structure.

So we can write hypothesis in compact form:

Logistic regression cost function and gradient descent

Simply speaking cost function is similar to linear regression cost function where linear hypothesis is replaced with logistic hypothesis. Without proof cost function can be represented as follows:

In order to minimize cost function we need to use gradient function:

We need to minimize J(Θ) we run gradient descent algorithm where each Θ is adjusted during several iterations until converge.

We can write a cost and gradient functions python code:

def cost(theta, X, y):
    '''  logistic regression cost    '''
    sxt = sigmoid(, theta));
    mcost = (-y)*np.log(sxt) - (1-y)*np.log(1-sxt)
    return mcost.mean()  
def gradient(theta, X, y):
    '''  logistic regression gradient    '''
    sxt = sigmoid(, theta))
    #label and prediction difference
    err = sxt - y 
    grad =, sxt) / y.size # gradient vector
    return grad

In order to minimize cost function we have two ways – write our own loop where we step by step update theta parameters until get to converging point, or use built in optimization routine which takes care of task. Since we have done manual optimization in linear regression, lets move on and use tools that make life easier. For optimization we are going to use function called fmin_bfgs which takes cost function, initial thetas and produce thetas of minimized function. Good news is that we even don’t need to prepare gradient function.

#create initial theta values
theta = 0.1* np.random.randn(n);
#use fmin_bfgs optimisation function find thetas
theta = opt.fmin_bfgs(cost, theta, args=(XX, Y));

After running minimization we get theta parameters that can be used in classification mode. Visually we can draw a decision boundary on our training data to see where data is separated in to classes.

Evaluating logistic regression classifier

We should not take results as granted. We need to test if model is valid and can be used for classification. We need to set up a test set with known labels. First of all we can write a predictor function which would produce class 0 if probability of hypothesis output is h<0.5 and class 1 if output is h>=0.5.

def predict(theta, X):
    #test new data
    m, n = X.shape
    p = np.zeros(shape=(m, 1));
    h = sigmoid(;
    p = 1*(h>=0.5);
    return p

Then we can prepare some test data to make sure if classifier works OK.

#test classifier with some random data
XXX = np.array([[1.0, 5.0, 1.4], #0
                [1.0, 1.4, 2.0], #0
                [1.0, 7.0, 1.0]]); #1
print(predict(theta, XXX));

And finally we can run our code on training data or separate validation set to see classification accuracy.

#test on labeled data
p = predict(theta, XX);                
print ('Accuracy: %f' % ((Y[np.where(p == Y)].size / float(Y.size)) * 100.0));

As we can see accuracy is 94%.

We can double check model on data by using WEKA software by selecting same data and Logistic classifier:

we can see that using same model in WEKA we get also 94% accuracy.

Next time we will touch more complicated case of Logistic regression which include regularization, multidimensional sets. As you start digging deeper, you will find that logistic regression is only one of many available classification tools.

If you would like to test algorithm by yourself, here is of python code and dataset.

Leave a Reply

Your email address will not be published. Required fields are marked *