Understanding unsupervised learning with clustering

Clustering is a statistical method that attempts to group the points in a dataset according to a distance measure, usually the Euclidean distance, which calculates the root of the squared differences between coordinates of a pair of points. To put this simply, those points that are classified within the same cluster are closer (in terms of the distance defined) to each other than they are to the points belonging to other clusters. At the same time, the larger the distance between two clusters, the better we can distinguish them. This is similar to saying that we try to build groups in which members are more alike and are more different to members of other groups.

It is clear that the most important part of a clustering algorithm is to define and calculate the distance between two given points and to iteratively assign the points to the defined clusters, until there is no change in the cluster composition. 

There are a few points to consider before trying a clustering analysis. Not every type of data is adequate for clustering. For example, we cannot use binary data since it is not possible to define distances. The values are either 1 or 0, and there is no value in-between. This excludes the type of data generated by one-hot encoding. Only data that shows some ordering or scale is useful for clustering. Even if the data values are real (such as, for example, a client's expenditure amounts or annual income), it is better to group them in a scale of ranges.

Some examples of clustering use cases are as follows:

• Automatic grouping of IT alerts to assign priorities and solve them accordingly
• Analysis of customer communication through different channels (segmentation in time periods) 
• Criminal profiling
• Urban mobility analysis
• Fraud detection (looking for outliers)
• Analysis of athletes' performances
• Crime analysis by geography
• Delivery logistics
• Classification of documents

Now, let's go through some examples that explains the concept of clustering algorithms.