Preprocessing: Scalers

1/**2 * Example 18: Preprocessing: Scalers3 *4 * Feature scaling puts columns on comparable ranges, which matters for5 * distance-based and gradient-based models. This example applies seven6 * scalers to the same small table that has one large outlier in each of7 * the first and last columns.8 *9 * Fit a scaler on the training data only, then reuse it with transform().10 */11 12import { type Tensor, tensor } from "deepbox/ndarray";13import {14  MaxAbsScaler,15  MinMaxScaler,16  Normalizer,17  PowerTransformer,18  QuantileTransformer,19  RobustScaler,20  StandardScaler,21} from "deepbox/preprocess";22 23console.log("=== Preprocessing: Scalers ===\n");24 25// First three rows, to keep the printed output short.26const first3 = (t: Tensor) => t.slice({ start: 0, end: 3 });27 28// Sample data with different scales29const X = tensor([30  [1, 100, 0.01],31  [2, 200, 0.02],32  [3, 300, 0.03],33  [4, 400, 0.04],34  [5, 500, 0.05],35  [100, 50, 0.5],36]);37 38// ---------------------------------------------------------------------------39// Part 1: StandardScaler: zero mean, unit variance40// ---------------------------------------------------------------------------41console.log("--- Part 1: StandardScaler ---");42 43const ss = new StandardScaler();44ss.fit(X);45const XStd = ss.transform(X);46console.log("Scaled (first 3 rows):\n", first3(XStd).toString());47console.log("Column means after scaling:", XStd.mean(0).toString());48 49const XInv = ss.inverseTransform(XStd);50console.log("Inverse transform recovers the first row:", first3(XInv).slice(0).toString());51 52// ---------------------------------------------------------------------------53// Part 2: MinMaxScaler: scale to [0, 1]54// ---------------------------------------------------------------------------55console.log("\n--- Part 2: MinMaxScaler ---");56 57const mms = new MinMaxScaler();58mms.fit(X);59const XMinMax = mms.transform(X);60console.log("Scaled (first 3 rows):\n", first3(XMinMax).toString());61 62// ---------------------------------------------------------------------------63// Part 3: RobustScaler: uses median and IQR (robust to outliers)64// ---------------------------------------------------------------------------65console.log("\n--- Part 3: RobustScaler ---");66 67const rs = new RobustScaler();68rs.fit(X);69const XRobust = rs.transform(X);70console.log("Scaled (first 3 rows):\n", first3(XRobust).toString());71 72// ---------------------------------------------------------------------------73// Part 4: MaxAbsScaler: scale by maximum absolute value74// ---------------------------------------------------------------------------75console.log("\n--- Part 4: MaxAbsScaler ---");76 77const mas = new MaxAbsScaler();78mas.fit(X);79const XMaxAbs = mas.transform(X);80console.log("Scaled (first 3 rows):\n", first3(XMaxAbs).toString());81 82// ---------------------------------------------------------------------------83// Part 5: Normalizer: normalize each sample (row) to unit norm84// ---------------------------------------------------------------------------85console.log("\n--- Part 5: Normalizer ---");86 87const norm = new Normalizer();88const XNorm = norm.transform(X);89console.log("Normalized (first 3 rows):\n", first3(XNorm).toString());90console.log("Row norms:", XNorm.square().sum(1).sqrt().toString());91 92// ---------------------------------------------------------------------------93// Part 6: PowerTransformer: Gaussian-like transformation94// ---------------------------------------------------------------------------95console.log("\n--- Part 6: PowerTransformer ---");96 97const pt = new PowerTransformer();98pt.fit(X);99const XPower = pt.transform(X);100console.log("Transformed (first 3 rows):\n", first3(XPower).toString());101 102// Deepbox's PowerTransformer does not standardize by default (scikit-learn does).103// Pass { standardize: true } to get zero mean and unit variance per column.104const ptStd = new PowerTransformer({ standardize: true });105console.log("Column means with standardize: true:", ptStd.fitTransform(X).mean(0).toString());106 107// ---------------------------------------------------------------------------108// Part 7: QuantileTransformer: map to uniform or normal distribution109// ---------------------------------------------------------------------------110console.log("\n--- Part 7: QuantileTransformer ---");111 112const qt = new QuantileTransformer();113qt.fit(X);114const XQuantile = qt.transform(X);115console.log("Transformed (first 3 rows):\n", first3(XQuantile).toString());116 117console.log("\n=== Preprocessing: Scalers Complete ===");