Determining Cluster Count Hyper Parameter for Unsupervised Learning KMEAN Models

Determining Cluster Counts with Elbow Method

Determining the optimal number of clusters in unsupervised machine learning, particularly for the KMeans algorithm, is a crucial step that significantly influences the algorithm's effectiveness and the meaningfulness of its results. KMeans works by partitioning data into clusters based on feature similarity, with the goal of minimizing variance within each cluster while maximizing variance between clusters. However, the algorithm requires the number of clusters to be specified a priori, and this is not always straightforward. Choosing too few clusters can lead to oversimplification, where distinct groups are improperly merged, while too many clusters might overfit the data, capturing noise and anomalies as separate groups. An optimal cluster count strikes a balance, ensuring that the model accurately captures the inherent structure of the data without overcomplicating it. This is essential for actionable insights and effective decision-making in various applications like market segmentation, anomaly detection, or organizing large data sets into understandable groups. Therefore, techniques like the elbow method (shown below), silhouette analysis, or the Davies–Bouldin index are often employed to estimate the most suitable number of clusters, enhancing the practical utility and interpretability of KMeans models.

Code

We stage proprietary crypto metrics accessible by Plus+ members (not shown) and then scale the data to be in a range of 0 or 1. KMEANs unsupervised learning uses numeric data to make unseen associations. If values are drastically different, it will negatively influence the algorithm. It is important to scale all of the data before running this clustering. Once scaling is done, the algorithm is run for increasing number of cluster counts. The average distance between the cluster and the data is called the inertia, and we run the algorithm for several cluster counts. The elbow graph is created in order to visualize where we see the rate of decrease in cluster counts decline significantly. The decline in the WCSS vs the cluster count indicates where increasing the number of clusters loses most predictive effect. Choose the cluster where the data seems to indicate that there is no more reduction in WCSS to cluster addition. Using the below elbow graph, we would choose 4 clusters. (Although choosing 5 or 6 clusters would also be reasonable based on the results).