UNPKG

clustering-tfjs

Version:

High-performance TypeScript clustering algorithms (K-Means, Spectral, Agglomerative) with TensorFlow.js acceleration and scikit-learn compatibility

139 lines (138 loc) 6.63 kB
"use strict"; var __createBinding = (this && this.__createBinding) || (Object.create ? (function(o, m, k, k2) { if (k2 === undefined) k2 = k; var desc = Object.getOwnPropertyDescriptor(m, k); if (!desc || ("get" in desc ? !m.__esModule : desc.writable || desc.configurable)) { desc = { enumerable: true, get: function() { return m[k]; } }; } Object.defineProperty(o, k2, desc); }) : (function(o, m, k, k2) { if (k2 === undefined) k2 = k; o[k2] = m[k]; })); var __setModuleDefault = (this && this.__setModuleDefault) || (Object.create ? (function(o, v) { Object.defineProperty(o, "default", { enumerable: true, value: v }); }) : function(o, v) { o["default"] = v; }); var __importStar = (this && this.__importStar) || (function () { var ownKeys = function(o) { ownKeys = Object.getOwnPropertyNames || function (o) { var ar = []; for (var k in o) if (Object.prototype.hasOwnProperty.call(o, k)) ar[ar.length] = k; return ar; }; return ownKeys(o); }; return function (mod) { if (mod && mod.__esModule) return mod; var result = {}; if (mod != null) for (var k = ownKeys(mod), i = 0; i < k.length; i++) if (k[i] !== "default") __createBinding(result, mod, k[i]); __setModuleDefault(result, mod); return result; }; })(); Object.defineProperty(exports, "__esModule", { value: true }); exports.SpectralClusteringConsensus = void 0; const tf = __importStar(require("../tf-adapter")); const spectral_1 = require("./spectral"); const tensor_utils_1 = require("../utils/tensor-utils"); /** * SpectralClustering with consensus clustering to improve robustness. * Runs k-means multiple times and takes majority vote for each point. */ class SpectralClusteringConsensus extends spectral_1.SpectralClustering { constructor(params) { super(params); this.consensusRuns = params.consensusRuns ?? 50; } async fit(X) { // We need a custom implementation that recomputes the embedding // because the parent class doesn't store it const Xtensor = (0, tensor_utils_1.isTensor)(X) ? tf.cast(X, 'float32') : tf.tensor2d(X, undefined, 'float32'); // Build affinity matrix (reuse parent logic) const computeAffinityMatrix = spectral_1.SpectralClustering.computeAffinityMatrix; this.affinityMatrix_ = computeAffinityMatrix(Xtensor, this.params); const affinitySum = (await this.affinityMatrix_.sum().data())[0]; if (affinitySum === 0) { throw new Error('Affinity matrix contains only zeros – cannot perform spectral clustering.'); } // Detect connected components const { detectConnectedComponents } = await Promise.resolve().then(() => __importStar(require('../utils/connected_components'))); const { numComponents, isFullyConnected, componentLabels } = detectConnectedComponents(this.affinityMatrix_); if (!isFullyConnected) { console.warn('Graph is not fully connected, spectral embedding may not work as expected.'); } let U; // If graph is disconnected and has enough components, use component indicators if (!isFullyConnected && numComponents >= this.params.nClusters) { const { createComponentIndicators } = await Promise.resolve().then(() => __importStar(require('../utils/component_indicators'))); U = createComponentIndicators(componentLabels, numComponents, numComponents); } else { // Standard approach: compute Laplacian and eigenvectors const { normalised_laplacian } = await Promise.resolve().then(() => __importStar(require('../utils/laplacian'))); const laplacian = tf.tidy(() => normalised_laplacian(this.affinityMatrix_)); const { smallest_eigenvectors_with_values } = await Promise.resolve().then(() => __importStar(require('../utils/smallest_eigenvectors_with_values'))); const numEigenvectors = Math.max(this.params.nClusters, numComponents); const { eigenvectors: U_full, eigenvalues } = smallest_eigenvectors_with_values(laplacian, numEigenvectors); // Apply diffusion map scaling const U_scaled = tf.tidy(() => { const numToUse = this.params.nClusters; const eigenvals = tf.slice(eigenvalues, [0], [numToUse]); const scalingFactors = tf.sqrt(tf.maximum(tf.scalar(0), tf.sub(tf.scalar(1), eigenvals))); const scalingFactors2D = scalingFactors.reshape([1, -1]); const U_selected = tf.slice(U_full, [0, 0], [-1, numToUse]); return U_selected.mul(scalingFactors2D); }); U = U_scaled; laplacian.dispose(); eigenvalues.dispose(); U_full.dispose(); } // Run k-means multiple times with different random seeds const { KMeans } = await Promise.resolve().then(() => __importStar(require('./kmeans'))); const allLabels = []; for (let run = 0; run < this.consensusRuns; run++) { const km = new KMeans({ nClusters: this.params.nClusters, randomState: (this.params.randomState ?? 42) + run, nInit: 1, // Single init per run, we handle multiple runs here }); await km.fit(U); allLabels.push(km.labels_); } // Consensus: for each point, take the most common label const n = allLabels[0].length; const consensusLabels = []; for (let i = 0; i < n; i++) { // Get all labels for point i const labelsForPoint = allLabels.map((labels) => labels[i]); // Count occurrences const counts = new Map(); for (const label of labelsForPoint) { counts.set(label, (counts.get(label) || 0) + 1); } // Find most common label let maxCount = 0; let consensusLabel = 0; for (const [label, count] of counts) { if (count > maxCount) { maxCount = count; consensusLabel = label; } } consensusLabels.push(consensusLabel); } // Update labels this.labels_ = consensusLabels; // Cleanup U.dispose(); if (!(0, tensor_utils_1.isTensor)(X)) { Xtensor.dispose(); } } } exports.SpectralClusteringConsensus = SpectralClusteringConsensus;