Skip to content
Deepbox
35advanced

Advanced Clustering

Runs eight clustering algorithms on synthetic data and scores them with the silhouette score and the adjusted Rand index. Seven run on well-separated blobs. Spectral clustering runs on two interleaved moons, a non-convex shape.

What you will learn

  • Use deepbox/ml for AgglomerativeClustering, GaussianMixture, SpectralClustering, OPTICS, MiniBatchKMeans, MeanShift, Birch, AffinityPropagation.
  • Use deepbox/metrics for silhouetteScore, adjustedRandScore.
  • Use deepbox/datasets for makeBlobs, makeMoons.
  • Runs eight clustering algorithms on synthetic data and scores them with the silhouette score and the adjusted Rand index. Seven run on well-separated blobs. Spectral clustering runs on two interleaved moons, a non-convex shape.

Source

index.ts
/** * Example 35: Advanced Clustering * * MiniBatchKMeans, AgglomerativeClustering, GaussianMixture, SpectralClustering, * OPTICS, MeanShift, Birch and AffinityPropagation, scored with the silhouette * score and the adjusted Rand index (ARI) on synthetic blobs and moons. */ import { makeBlobs, makeMoons } from "deepbox/datasets";import { adjustedRandScore, silhouetteScore } from "deepbox/metrics";import {  AffinityPropagation,  AgglomerativeClustering,  Birch,  GaussianMixture,  KMeans,  MeanShift,  MiniBatchKMeans,  OPTICS,  SpectralClustering,} from "deepbox/ml"; console.log("=".repeat(60));console.log("Example 35: Advanced Clustering");console.log("=".repeat(60)); // ============================================================================// Generate synthetic datasets// ============================================================================ // Well-separated blobs for standard clusteringconst [XBlobs, yBlobs] = makeBlobs({  nSamples: 200,  centers: 4,  clusterStd: 0.8,  randomState: 42,}); // Non-convex moons for advanced methodsconst [XMoons, yMoons] = makeMoons({  nSamples: 200,  noise: 0.1,  randomState: 42,}); console.log(`Blobs dataset: ${XBlobs.shape[0]} samples, ${XBlobs.shape[1]} features, 4 clusters`);console.log(`Moons dataset: ${XMoons.shape[0]} samples, ${XMoons.shape[1]} features, 2 clusters`); // ============================================================================// Part 1: MiniBatchKMeans// ============================================================================console.log("\nPart 1: MiniBatchKMeans");console.log("-".repeat(60)); // MiniBatchKMeans updates the centers from small random batches instead of the full datasetconst mbkmeans = new MiniBatchKMeans({  nClusters: 4,  batchSize: 50,  maxIter: 100,  randomState: 42,});mbkmeans.fit(XBlobs); const mbkLabels = mbkmeans.predict(XBlobs);const mbkSil = silhouetteScore(XBlobs, mbkLabels);const mbkAri = adjustedRandScore(yBlobs, mbkLabels); console.log("MiniBatchKMeans (k=4, batchSize=50):");console.log(`  Silhouette Score:     ${mbkSil.toFixed(4)}`);console.log(`  Adjusted Rand Index:  ${mbkAri.toFixed(4)}`);console.log(`  Cluster centers shape: ${mbkmeans.clusterCenters.shape}`); // Compare with standard KMeansconst kmeans = new KMeans({ nClusters: 4, randomState: 42 });kmeans.fit(XBlobs);const kmLabels = kmeans.predict(XBlobs);const kmSil = silhouetteScore(XBlobs, kmLabels);console.log(`  KMeans Silhouette (for comparison): ${kmSil.toFixed(4)}`); // ============================================================================// Part 2: Agglomerative Clustering// ============================================================================console.log("\nPart 2: Agglomerative Clustering");console.log("-".repeat(60)); // Agglomerative clustering starts with one cluster per sample and repeatedly merges the closest pairconst aggWard = new AgglomerativeClustering({  nClusters: 4,  linkage: "ward",});aggWard.fit(XBlobs); const aggLabels = aggWard.labels;const aggSil = silhouetteScore(XBlobs, aggLabels);const aggAri = adjustedRandScore(yBlobs, aggLabels); console.log("Agglomerative (Ward linkage, k=4):");console.log(`  Silhouette Score:    ${aggSil.toFixed(4)}`);console.log(`  Adjusted Rand Index: ${aggAri.toFixed(4)}`); // Try different linkagesfor (const linkage of ["complete", "average", "single"] as const) {  const agg = new AgglomerativeClustering({ nClusters: 4, linkage });  agg.fit(XBlobs);  const sil = silhouetteScore(XBlobs, agg.labels);  console.log(`  ${linkage.padEnd(10)} linkage, Silhouette: ${sil.toFixed(4)}`);} // ============================================================================// Part 3: Gaussian Mixture Model// ============================================================================console.log("\nPart 3: Gaussian Mixture Model");console.log("-".repeat(60)); // A Gaussian mixture models the data as a weighted sum of Gaussian distributionsconst gmm = new GaussianMixture({  nComponents: 4,  maxIter: 100,});gmm.fit(XBlobs); const gmmLabels = gmm.predict(XBlobs);const gmmSil = silhouetteScore(XBlobs, gmmLabels);const gmmAri = adjustedRandScore(yBlobs, gmmLabels); console.log("Gaussian Mixture (4 components):");console.log(`  Silhouette Score:    ${gmmSil.toFixed(4)}`);console.log(`  Adjusted Rand Index: ${gmmAri.toFixed(4)}`); // ============================================================================// Part 4: Spectral Clustering// ============================================================================console.log("\nPart 4: Spectral Clustering");console.log("-".repeat(60)); // Spectral clustering uses the eigenvectors of a similarity matrix, which suits non-convex shapesconst spectral = new SpectralClustering({  nClusters: 2,  affinity: "rbf",  gamma: 10,});spectral.fit(XMoons); const specLabels = spectral.labels;const specSil = silhouetteScore(XMoons, specLabels);const specAri = adjustedRandScore(yMoons, specLabels); console.log("Spectral Clustering on Moons (k=2, RBF kernel):");console.log(`  Silhouette Score:    ${specSil.toFixed(4)}`);console.log(`  Adjusted Rand Index: ${specAri.toFixed(4)}`); // ============================================================================// Part 5: OPTICS// ============================================================================console.log("\nPart 5: OPTICS");console.log("-".repeat(60)); // OPTICS is density-based and does not need a fixed neighborhood radius (eps)const optics = new OPTICS({  minSamples: 5,});optics.fit(XBlobs); const optLabels = optics.labels;const optAri = adjustedRandScore(yBlobs, optLabels); console.log("OPTICS (minSamples=5):");console.log(`  Adjusted Rand Index: ${optAri.toFixed(4)}`);console.log(`  Reachability values: ${optics.reachability.length}`); // ============================================================================// Part 6: MeanShift// ============================================================================console.log("\nPart 6: MeanShift");console.log("-".repeat(60)); // MeanShift moves points uphill to the modes of the density estimate. Each mode is one cluster.const meanshift = new MeanShift({  bandwidth: "auto",});meanshift.fit(XBlobs); const msLabels = meanshift.labels;const msSil = silhouetteScore(XBlobs, msLabels); console.log("MeanShift (auto bandwidth):");console.log(`  Silhouette Score:    ${msSil.toFixed(4)}`);console.log(`  Cluster centers shape: ${meanshift.clusterCenters.shape}`); // ============================================================================// Part 7: Birch// ============================================================================console.log("\nPart 7: Birch");console.log("-".repeat(60)); // Birch summarizes the data in a clustering-feature tree, so it needs one pass over the samplesconst birch = new Birch({  nClusters: 4,  threshold: 0.5,  branchingFactor: 50,});birch.fit(XBlobs); const birchLabels = birch.labels;const birchSil = silhouetteScore(XBlobs, birchLabels);const birchAri = adjustedRandScore(yBlobs, birchLabels); console.log("Birch (k=4, threshold=0.5):");console.log(`  Silhouette Score:    ${birchSil.toFixed(4)}`);console.log(`  Adjusted Rand Index: ${birchAri.toFixed(4)}`); // ============================================================================// Part 8: Affinity Propagation// ============================================================================console.log("\nPart 8: Affinity Propagation");console.log("-".repeat(60)); // Affinity propagation passes messages between samples and chooses the number of clusters itselfconst ap = new AffinityPropagation({  damping: 0.9,  maxIter: 200,});ap.fit(XBlobs); const apLabels = ap.labels;const apSil = silhouetteScore(XBlobs, apLabels); console.log("Affinity Propagation (damping=0.9):");console.log(`  Silhouette Score:    ${apSil.toFixed(4)}`);console.log(`  Cluster centers shape: ${ap.clusterCenters.shape}`); // ============================================================================// Summary Comparison// ============================================================================console.log("\nClustering Comparison on Blobs Dataset");console.log("-".repeat(60)); console.log("┌─────────────────────────┬────────────┬──────────┐");console.log("│ Algorithm               │ Silhouette │ ARI      │");console.log("├─────────────────────────┼────────────┼──────────┤"); const comparisons = [  { name: "KMeans", sil: kmSil, ari: adjustedRandScore(yBlobs, kmLabels) },  { name: "MiniBatchKMeans", sil: mbkSil, ari: mbkAri },  { name: "Agglomerative (Ward)", sil: aggSil, ari: aggAri },  { name: "Gaussian Mixture", sil: gmmSil, ari: gmmAri },  { name: "MeanShift", sil: msSil, ari: adjustedRandScore(yBlobs, msLabels) },  { name: "Birch", sil: birchSil, ari: birchAri },  { name: "Affinity Propagation", sil: apSil, ari: adjustedRandScore(yBlobs, apLabels) },]; for (const c of comparisons) {  const name = c.name.padEnd(23);  const sil = c.sil.toFixed(4).padStart(10);  const ari = c.ari.toFixed(4).padStart(8);  console.log(`│ ${name} │ ${sil} │ ${ari} │`);}console.log("└─────────────────────────┴────────────┴──────────┘"); // ============================================================================// Summary// ============================================================================console.log("\nKey Takeaways");console.log("-".repeat(60));console.log("• MiniBatchKMeans: KMeans on mini-batches, cheaper for large datasets");console.log("• Agglomerative: hierarchical clustering with a choice of linkage");console.log("• Gaussian Mixture: probabilistic assignments instead of hard ones");console.log("• Spectral: graph-based, handles non-convex shapes such as moons");console.log("• OPTICS: density-based, no eps to choose up front");console.log("• MeanShift: finds the number of clusters from the density modes");console.log("• Birch: one pass over the data, low memory");console.log("• Affinity Propagation: finds the number of clusters by message passing"); console.log("\nAdvanced Clustering Example Complete!");console.log("=".repeat(60));

Output

Console output
Console output only: a silhouette score and ARI per algorithm, and a comparison table for the blobs dataset.