In supervised learning, models are trained on labeled data. They learn to predict answers that are already known. However, in practice, we may have unlabeled data without any predefined answers. In this case, we use unsupervised learning.

Unsupervised learning is a type of machine learning, where a model learns from unlabeled data and discovers patterns, relations, structure, or similarities in data by itself. The model explores the data, looking for some valuable information. Let us take the example of an online shop with thousands of customers. The shop collects data about their purchases, spending, and browsing history, but does not know anything about predefined customer categories. An unsupervised learning model will help the shop group customers with similar behavior.

There are different techniques of unsupervised learning, but the most popular categories are clustering, dimensionality reduction, and association rule learning.

With clustering, data points with similar characteristics are grouped. A model identifies similarities within data without any information about groups that should exist in data. For example, a company may want to divide customers into several groups based on their purchasing habits*. K-Means, DBSCAN, and Hierarchical Clustering* are the most popular clustering algorithms.

Real-life datasets may contain hundreds or even thousands of features. Having so many features makes data more difficult to visualize, process, and analyze. Dimensionality reduction techniques reduce the number of features in data and keep as much useful information as possible. For example, a dataset with hundreds of features can be reduced to two dimensions to visualize its patterns. PCA and t-SNE are the most common dimensionality reduction techniques.

Association rule learning detects relations between variables in a dataset. It is usually used to discover co-occurrence of items/events. For example, a supermarket can detect that customers who purchased bread and butter also purchased milk. Such information may help the business place products in a better way or make product recommendations. The most common algorithms are Apriori and FP-Growth.

The process starts with gathering and preparing unlabeled data. The algorithm analyzes data using its learning approach and searches for similarities, differences, relations, and patterns. Then the results are evaluated to see if the structure found is useful for solving the problem.

Unlike supervised learning, in unsupervised learning there are no predefined labels to measure all predictions directly. Therefore, evaluation of unsupervised learning algorithms usually requires different approaches like clustering metrics, visualization, or knowledge about the domain.

Unsupervised learning is needed when labeled data is not available, too expensive to gather, or not enough for data analysis. Unsupervised learning helps to uncover patterns that would be missed otherwise and may be useful for customer segmentation, anomaly detection, data exploration, and creating recommendation systems.

However, finding a pattern does not mean that it is useful. The results of unsupervised learning also should be interpreted and verified according to the problem. Understanding unsupervised learning is another important milestone in learning machine learning. We see that sometimes models do not need predefined answers to learn something new.