Automatic Cluster Discovery

One of the biggest challenges in unsupervised learning is choosing the number of clusters ( ).

While standard K-Means requires you to specify upfront, X-Means (available in DruidJS) automatically discovers the optimal number of clusters using the Bayesian Information Criterion (BIC).

Discovering clusters automatically...

How it works

1. Initial K: X-Means starts with a minimum number of clusters (typically ).
2. Recursive Splitting: For each cluster, the algorithm attempts to split it into two sub-clusters using K-Means.
3. Statistical Validation: It compares the BIC score of the original cluster versus the split. If the split provides a significantly better fit, the new clusters are kept.
4. Iterative Refinement: This process repeats until it reaches a maximum or no further improvements are found.

Benefits

• Zero Configuration: No need to run "elbow plots" or manual searches for .
• Objective Fit: Uses a robust statistical measure (BIC) to prevent over-fitting.
• Speed: Built on top of K-Means, it remains efficient for large datasets.

How-to (Code)

Use XMeans when you want the algorithm to determine the cluster count for you. You only need to provide a search range (K_min and K_max).

import * as druid from "@saehrimnir/druidjs";

const data = [
  /* ... high-dimensional data ... */
];

// Initialize XMeans with a search range for clusters
const xmeans = new druid.XMeans(data, {
  K_min: 2,
  K_max: 10,
});

// Run the discovery algorithm
// It returns an array of cluster indices for each point
const clusters = xmeans.get_cluster_list();

// You can check how many clusters were found
console.log(`Optimal number of clusters: ${xmeans.k}`);