What is Unsupervised Learning?

Unsupervised learning methods, (opposite to supervised learning methods) do not require any labels or responses along with the training data. Unsupervised ML algorithms learn patterns and relationships from the given raw data. Although these methods may have lots of uncertainties in the result, we can always obtain some useful information such as all kinds of unknown patterns in the data, and the useful features for categorization.

Unsupervised algorithms are called unsupervised because the machine learning model learns from data samples where the output is not known in advance.

Let’s make it clearer, suppose we have,

Input variables − m

But as opposed to supervised learning methods, there would be no corresponding output variable and for learning these unsupervised algorithms need to discover the interesting patterns in raw data.

Some of the well-known unsupervised machine learning algorithms are Hebbian Learning, K-means Clustering and Hierarchical Clustering.

Based on machine learning based tasks, we can divide unsupervised learning algorithms in following classes −

  • Clustering − Clustering is one among the most useful unsupervised ML methods. It is used to find the relationship patterns and similarity among the input data samples. After finding these patterns, unsupervised algorithm clusters the data samples having similarities into groups as illustrated in the diagram below −

  • Association − Association is used to find the related patterns that represent the interesting relationships between a variety of data items in a large dataset. One of the best examples of association is analyzing customer shopping patterns. This example is also known as Market Basket Analysis.

  • Dimensionality Reduction − Dimensionality Reduction, as name entails, is used to slash the number of feature variables. To do this, it selects a principal feature for every data sample. The main reason behind using this method is to eliminate the problem of feature space complexity also called curse of dimensionality. Principal Component Analysis (PCA) is one of the most popular dimensionality reduction methods.

  • Anomaly Detection − Anomaly detection, as name entails, finds the occurrences of anomalous observations i.e., the observations that do not occur generally. It uses learned knowledge to differentiate between normal and anomalous data points. Mean-shift clustering, and k-means clustering are some unsupervised ML algorithms that detect the anomalous data based on data features.