Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                
Scribd
Upload
51 views

Logistic Regression

Logistic regression models the probability of a binary outcome using a logistic function. It extends linear regression by applying the logistic function to the linear combination of features to constrain the output between 0 and 1. This allows the output to be interpreted as a probability. The logistic regression hypothesis represents the log odds of the probability of an output being 1. Regularization is added to the cost function to address overfitting by penalizing models with large coefficients.

Uploaded by

Michael Oluwakayode Adedeji
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
51 views

Logistic Regression

Logistic regression models the probability of a binary outcome using a logistic function. It extends linear regression by applying the logistic function to the linear combination of features to constrain the output between 0 and 1. This allows the output to be interpreted as a probability. The logistic regression hypothesis represents the log odds of the probability of an output being 1. Regularization is added to the cost function to address overfitting by penalizing models with large coefficients.

Uploaded by

Michael Oluwakayode Adedeji
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 37

Logistic Regression and

Regularization
Regression vs Classification
Having worked on regression problems for a while, we now move on to
classification problems.

Regression Problems => Real valued output in continuous range

Classification Problems => Discrete valued output in categories


Binary Classification
In binary classification, there are only 2 categories or classes, and This classes
are usually represented as 0 or 1.

The model is therefore expected to produce an output value that is either 0 or 1.

When the number of classes is more than 2 the classification problem is termed
Multiclass classification.
y ∊ {0,1}

In binary classification, often times depending on the nature of the problem;

0 is taken as the ‘negative class’ and 1 the ‘positive class’, or


0 is taken as the ‘false class’ and 1 the ‘true class’, or
0 is taken as the ‘no class’ and 1 the ‘yes class’

You are free to use any representation whatsoever.


How do we tackle the binary classification problem ?
We all know how linear regression works right ?

Let’s use our knowledge of linear regression and build an intuition from that, taking
note that in this case, our output is either 0 or 1.
Hypothesis Representation

Since the target label is either 0 or 1, our hypothesis has to lie in that range;
0 ≤ hθ(x) ≤ 1 (this is what we want)

Linear regression hypothesis is gives as


hθ(х) = θ0 + θ1x1+ θ2x2 + … + θnxn = θTx

And it varies as;


-∞ ≤ hθ(x) ≤ -∞

So, how do we achieve this; 0 ≤ hθ(x) ≤ 1


We can do this by passing θTx into a function, g(z) called ‘Sigmoid function’ or
‘Logistic function’

This function takes in whatever input you pass into it and produces an output
between 0 and 1.
This is now the hypothesis for logistic regression

And,
The hypothesis is ranges between 0 and 1, and can be modelled as the probability
that an output is 1.

E.g, hθ(x) = 0.7 is a 70% chance that our output is 1.

Therefore,

And,
Decision Boundary
The hypothesis gives values ranging from 0 to 1. In order to obtain the class of the
prediction, we can establish a threshold in which values greater than the threshold
are rounded up as 1 and lesser values are 0. We can choose 0.5
The logistic/sigmoid function has a very interesting behaviour. When its input is
greater than or equal to 0, its output is greater than or equal to 0.5, and when its
input is less than 0, its output is less than 0.5

The decision boundary is a line that separates the areas y=1 and y=0
Cost Function
What cost function should we use ?

Should we use MSE as in Linear regression ?


If our hypothesis is far from y, the cost tends to infinity.
If our hypothesis is exactly the same as y, the cost is zero.
Simplified Cost Function and Gradient Descent
We can write the cost function’s conditional cases in a more compact form
Vectorized implementation
Gradient Descent
The general form of gradient descent is:
Working out the partial derivative, we get,
And a vectorized implementation gives,
Multiclass Classification: One-vs-all
y ∊ {0,1, … , n}
Softmax Regression
Softmax function:
Regularization
Regularization is designed to address the problem of overfitting.
Causes of Overfitting:

Too many features


Addressing overfitting:
Cost Function
Optimization objective:
Regularized Linear Regression
Gradient Descent:
Normal Equation:
Regularized Logistic Regression
Cost function:
The cost function is regularized by adding a term to the end
Gradient Descent:

You might also like