Logistic Regression

Let’s enter the vast world of classification algorithms in machine learning.

In classification, the outcomes - the targets - are no more real-valued but class labels that can be associated with discrete numericals. While it could appear as a sub-case of the linear regression, we will see that we need to redefine our mathematical tools.

In this lecture we will cover a method for classifying data called logistic regression. The learning objectives are:

• Understand the necessity of the sigmoid function

• Understand the requirements for the cost function in the context of logistic regression

• Understand the role and structure of the cross-entropy function

• Write the pseudo-code of the gradient descent algorithm for classification

• Understand strategies for multiclass classification, including one-vs-all and softmax (or one-vs-one) approaches

Particle physics florishes of classification implementations. We will put the logistic regression in practice during a tutorial to classify collision data!