Clustering is a set of methods or algorithms that are used to find natural groupings according to predefined properties of variables in a dataset. The Merriam-Webster dictionary defines a cluster as "a number of similar things that occur together." Clustering in unsupervised learning is exactly what it means in the traditional sense. For example, how do you identify a bunch of grapes from far away? You have an intuitive sense without looking closely at the bunch whether the grapes are connected to each other or not. Clustering is just like that. An example of clustering is presented here:
Figure 1.2: A representation of two clusters in a dataset
In the preceding graph, the data points have two properties: cholesterol and blood pressure. The data points are classified into two clusters, or two bunches, according to the Euclidean distance between them. One cluster contains people who are clearly at high risk of heart disease and the other cluster contains people who are at low risk of heart disease. There can be more than two clusters, too, as in the following example:
Figure 1.3: A representation of three clusters in a dataset
In the preceding graph, there are three clusters. One additional group of people has high blood pressure but with low cholesterol. This group may or may not have a risk of heart disease. In further sections, clustering will be illustrated on real datasets in which the x and y coordinates denote actual quantities.
Like all methods of unsupervised learning, clustering is mostly used when we don't have labeled data – data with predefined classes – for training our models. Clustering uses various properties, such as Euclidean distance and Manhattan distance, to find patterns in the data and classify them according to similarities in their properties without having any labels for training. So, clustering has many use cases in fields where labeled data is unavailable or we want to find patterns that are not defined by labels.
The following are some applications of clustering:
Exploratory data analysis: When we have unlabeled data, we often do clustering to explore the underlying structure and categories of the dataset. For example, a retail store might want to explore how many different segments of customers they have, based on purchase history.
Generating training data: Sometimes, after processing unlabeled data with clustering methods, it can be labeled for further training with supervised learning algorithms. For example, two different classes that are unlabeled might form two entirely different clusters, and using their clusters, we can label data for further supervised learning algorithms that are more efficient in real-time classification than our unsupervised learning algorithms.
Recommender systems: With the help of clustering, we can find the properties of similar items and use these properties to make recommendations. For example, an e-commerce website, after finding customers in the same clusters, can recommend items to customers in that cluster based upon the items bought by other customers in that cluster.
Natural language processing: Clustering can be used for the grouping of similar words, texts, articles, or tweets, without labeled data. For example, you might want to group articles on the same topic automatically.
Anomaly detection: You can use clustering to find outliers. We're going to learn about this in Chapter 6, Anomaly Detection. Anomaly detection can also be used in cases where we have unbalanced classes in data, such as in the case of the detection of fraudulent credit card transactions.