Vincent and Gill explain supervised learning, where labeled data trains models to predict outcomes, versus unsupervised learning like clustering, which groups similar data without labels. Unsupervised models are harder to evaluate because there's no ground truth to check against. Semi-supervised learning splits the difference, using a few labeled examples plus clustering to fill in the rest.
Chapters
Hi there, my name is Vincent. Together with Gill, I will guide you through your first steps in the world of machine learning. In a previous video, you learned about three machine learning, classification, regression and clustering. As you might have felt, there are quite some similarities between classification and regression. For both, we try to find a function or a model which can later be used to predict labels or values for unseen observations. It is important that during the training of the function, labeled observations are available to the algorithm. We call these techniques supervised learning. Labeling can be a tedious work and it is often done by humans. There are other techniques which don't require labeled observations. These techniques are called unsupervised learning. You've already acquainted yourself with one of these techniques in a previous video, namely clustering. CL Clustering will find groups of observations that are similar and thus does not require specific labeled observations. In the next chapter, we'll talk about assessing the performance of your trained model. In supervised learning, we can use the real labels of observations and compare them with the labels we predicted. It's quite straightforward that you want your model's predictions to be as similar as possible to the real labels. With unsupervised learning, however, measuring the performance gets more difficult. We don't have any real labels to compare anything to. You'll learn some neat techniques to assess the quality of a clustering in the next chapter. As you get more experienced as a data scientist, you might notice that things aren't always black and white in machine learning. Some techniques overlap between supervised and unsupervised learning. With semi supervised learning, for example, you can have a lot of observations which are not labeled and a few which are. You can then first perform clustering to group all observations which are similar. Afterwards, you can use information about the clusters and about a few labeled observations to assign a class to unlabeled observations. This will give you more labeled observations to perform supervised learning on. Enough talking. Let's do some exercises.
