Model Evaluation Metrics

1/**2 * Example 24: Model Evaluation Metrics3 *4 * Metrics for classification, regression and clustering. Every metric5 * returns a plain number (or an array of numbers when asked for per-class6 * values), so results can be printed or compared directly.7 */8 9import {10  // Classification metrics11  accuracy,12  confusionMatrix,13  f1Score,14  jaccardScore,15  logLoss,16  mae,17  mape,18  matthewsCorrcoef,19  meanAbsolutePercentageError,20  mse,21  precision,22  // Regression metrics23  r2Score,24  recall,25  rmse,26  rocAucScore,27  // Clustering metrics28  silhouetteScore,29} from "deepbox/metrics";30import { tensor } from "deepbox/ndarray";31 32console.log("=== Model Evaluation Metrics ===\n");33 34// Binary classification metrics35console.log("1. Binary Classification Metrics:");36console.log("-".repeat(50));37 38const yTrueClass = tensor([1, 0, 1, 1, 0, 1, 0, 0, 1, 1]);39const yPredClass = tensor([1, 0, 1, 0, 0, 1, 0, 1, 1, 1]);40 41console.log("True labels:", yTrueClass.toString());42console.log("Predictions:", `${yPredClass.toString()}\n`);43 44const acc = accuracy(yTrueClass, yPredClass);45const prec = precision(yTrueClass, yPredClass);46const rec = recall(yTrueClass, yPredClass);47const f1 = f1Score(yTrueClass, yPredClass);48 49console.log(`Accuracy:  ${(acc * 100).toFixed(2)}%`);50console.log(`Precision: ${(prec * 100).toFixed(2)}%`);51console.log(`Recall:    ${(rec * 100).toFixed(2)}%`);52console.log(`F1-Score:  ${(f1 * 100).toFixed(2)}%`);53console.log(`Jaccard:   ${(jaccardScore(yTrueClass, yPredClass) * 100).toFixed(2)}%`);54console.log(`Matthews correlation: ${matthewsCorrcoef(yTrueClass, yPredClass).toFixed(4)}\n`);55 56const cm = confusionMatrix(yTrueClass, yPredClass);57console.log("Confusion Matrix:");58console.log(cm.toString());59console.log("Rows are true classes, columns are predicted classes: [[TN, FP], [FN, TP]]\n");60 61// Probability-based metrics need scores, not labels.62console.log("2. Probability-Based Metrics (binary):");63console.log("-".repeat(50));64 65const yProba = tensor([0.9, 0.2, 0.8, 0.4, 0.1, 0.7, 0.3, 0.6, 0.85, 0.95]);66console.log(`ROC AUC:  ${rocAucScore(yTrueClass, yProba).toFixed(4)}`);67console.log(`Log loss: ${logLoss(yTrueClass, yProba).toFixed(4)}\n`);68 69// Multiclass metrics: choose how per-class scores are combined with `average`.70console.log("3. Multiclass Metrics:");71console.log("-".repeat(50));72 73const yTrueMulti = tensor([0, 1, 2, 2, 1, 0, 2, 1, 0, 2]);74const yPredMulti = tensor([0, 2, 2, 2, 1, 0, 1, 1, 0, 2]);75 76for (const average of ["macro", "micro", "weighted"] as const) {77  const p = precision(yTrueMulti, yPredMulti, { average });78  const r = recall(yTrueMulti, yPredMulti, { average });79  const f = f1Score(yTrueMulti, yPredMulti, { average });80  console.log(81    `${average.padEnd(9)} precision=${p.toFixed(4)} recall=${r.toFixed(4)} f1=${f.toFixed(4)}`82  );83}84 85// average: null returns one value per class, in label order.86const perClass = f1Score(yTrueMulti, yPredMulti, { average: null });87console.log(`Per-class F1: ${perClass.map((v) => v.toFixed(4)).join(", ")}`);88 89// Multiclass ROC AUC takes an [n, classes] matrix of probabilities that sum to 1 per row.90const yProbaMulti = tensor([91  [0.8, 0.1, 0.1],92  [0.2, 0.3, 0.5],93  [0.1, 0.2, 0.7],94  [0.1, 0.1, 0.8],95  [0.2, 0.6, 0.2],96  [0.7, 0.2, 0.1],97  [0.1, 0.5, 0.4],98  [0.3, 0.5, 0.2],99  [0.6, 0.3, 0.1],100  [0.2, 0.2, 0.6],101]);102console.log(`Multiclass ROC AUC (macro): ${rocAucScore(yTrueMulti, yProbaMulti).toFixed(4)}`);103console.log(`Multiclass log loss: ${logLoss(yTrueMulti, yProbaMulti).toFixed(4)}\n`);104 105// Regression metrics106console.log("4. Regression Metrics:");107console.log("-".repeat(50));108 109const yTrueReg = tensor([3.0, 0.5, 2.0, 7.0, 4.2]);110const yPredReg = tensor([2.5, 0.6, 2.1, 7.8, 4.0]);111 112console.log("True values:", yTrueReg.toString());113console.log("Predictions:", `${yPredReg.toString()}\n`);114 115const r2 = r2Score(yTrueReg, yPredReg);116const mseVal = mse(yTrueReg, yPredReg);117const rmseVal = rmse(yTrueReg, yPredReg);118const maeVal = mae(yTrueReg, yPredReg);119 120console.log(`R² Score: ${r2.toFixed(4)}`);121console.log(`MSE:      ${mseVal.toFixed(4)}`);122console.log(`RMSE:     ${rmseVal.toFixed(4)}`);123console.log(`MAE:      ${maeVal.toFixed(4)}`);124 125// mape already returns a percentage. scikit-learn's version returns a fraction,126// which is what meanAbsolutePercentageError returns.127console.log(`MAPE:     ${mape(yTrueReg, yPredReg).toFixed(2)}% (mape returns a percentage)`);128console.log(129  `meanAbsolutePercentageError: ${meanAbsolutePercentageError(yTrueReg, yPredReg).toFixed(4)} (a fraction)\n`130);131 132// Clustering metrics133console.log("5. Clustering Metrics:");134console.log("-".repeat(50));135 136const XCluster = tensor([137  [1, 2],138  [1.5, 1.8],139  [5, 8],140  [8, 8],141  [1, 0.6],142  [9, 11],143]);144const labels = tensor([0, 0, 1, 1, 0, 1]);145 146const silhouette = silhouetteScore(XCluster, labels);147console.log(`Silhouette Score: ${silhouette.toFixed(4)}`);148console.log("Range: [-1, 1], higher is better");149console.log("Measures how close points are to their own cluster compared with the next one\n");150 151console.log("Metric selection guide:");152console.log(153  "  Classification: F1 for imbalanced data, ROC AUC and log loss when you have probabilities"154);155console.log("  Regression: R² for variance explained, MAE for an error in the target's units");156console.log("  Clustering: silhouette for cluster separation");