Thread by @svpino

Multi-label classification problems seem to get less attention than binary or multi-class classification problems.

They are widespread in real life, so you should definitely know how to recognize them and solve them.

This is a short 🧵👇

In machine learning, you are in front of a multi-label classification problem whenever you want to classify your samples using one or more labels.

For example, you could classify a movie as "Horror," "Thriller," and "Classic" simultaneously.

[2 / 5]

In more formal terms:

Multi-label classification is a predictive modeling technique that predicts zero or more mutually non-exclusive labels.

[3 / 5]

You can use neural networks to tackle multi-label classification problems:

▫️ Set output layer size to the number of labels
▫️ Use sigmoid activation in the output layer
▫️ Use binary cross-entropy loss

[3 / 5]

Because we are using a sigmoid activation function, a prediction will be an array of values between 0 and 1. For example: [0.9943, 0.2343, 0.7592]

If we round the result, we will get: [1, 0, 1]

This indicates that labels 0 and 2 should be assigned to the input.

[5 / 5]