Clustering is an important non-supervised learning technique. It aims to split data into certain clusters. For instance, if you input data pertaining to shoppers in the local grocery market, clustering that could output age based clusters of say < 12 years, 12-18, 18-60, and > 60. Another intuitive example is banking, clustering financial data for a large group of individuals could output income-based clusters, say 3 pertaining to the lower middle, upper middle, and upper classes.
The most basic and intuitive method of clustering is K-means, which identifies K clusters. It randomly initialises K centroids, marks the points near it, and repeats until it repeats K averaged out cluster centroids. Hierarchal analysis, however, is based on a much different principle.
There are two methods of hierarchal analysis: agglomerative and divisive. Agglomerative puts each data point into a cluster of its own. Hence, if you input 6000 points, you start out with 600 clusters. It then clusters the closest points (and then clusters) and repeats this process until there is only one giant cluster encompassing the entire dataset left. Divisive does the exact opposite. It starts with one giant clusters, and splits each cluster (repeatedly) until each point is its own cluster. The results are plotted on a special plot, called a dendrogram. The longest vertical line segment on the dendrogram gets to be the optimum number of clusters for analysis. This will be much easier to understand when I show a dendrogram below.
Note down the libraries I’ve imported. The dataset is fairly straightforward. You have an ID for each customer, and financial data corresponding to that ID. You’ll notice that the standard deviation or range for features is quite different. Where balance frequency tends to stay close to 1 for each ID, account balance is wildly different. This can cause issues during clustering, so that’s why i’ve scaled my data so that each feature is similar to each other feature in relative terms.
Here is the dendrogram for the data. The y-axis represents the ‘closeness’ of each individual data-point/cluster. You’d obviously expect y to be maxed out when there’s only 1 cluster, so that’s no surprise. Now, looking at this graph, we must select the number of clusters for our model. A general rule of thumb here is to take the number of clusters pertaining to the longest vertical line visible here. The distance of each vertical line from each other represents how faraway those clusters are. Hence, you want a small number of clusters (not necessary, but in this application, optimal), but also want your clusters to be spaced far apart, so that they clearly represent different groups of people (in this context).
I’m taking 3 clusters, which corresponds to a vertical axis value of 23. 3 clusters also intuitively makes sense to me as any customer can broadly be classified into lower, middle, and upper class. Of course, there are subdivisions inside these 3 broad categories too, and you might argue that the lower class wouldn’t even be represented here, so we can say that these 3 clusters correspond to the lower middle, upper middle, and upper classes.
Here is a diagrammatic representation of what I’ve chosen.
After building the model, all that’s left is visualising the results. There are more than two features, so I’m arbitrarily selecting two, plotting all points using those features for my axes, and giving each point a color that corresponds to all other points in its cluster.
You’ll notice that there is a lot of data here, but also a clear pattern. Points belonging to the purple cluster visibly tend towards the upper left corner. Similarly, points in the teal cluster tend to the bottom left corner, and points in the yellow cluster tend to the bottom right corner.