Skip to content
Deepbox
38intermediate

Feature Engineering & Preprocessing

Nine short parts on preprocessing: imputation, feature selection, text vectorizers, polynomial and spline features, median-based and power scalers, and cross-validation splitters. PowerTransformer has standardize set to false by default, while scikit-learn uses true. The example shows both settings.

What you will learn

  • Use deepbox/preprocess for SimpleImputer, KNNImputer, SelectKBest, fClassif, VarianceThreshold, TfidfVectorizer, CountVectorizer, SplineTransformer, PolynomialFeatures, RobustScaler, PowerTransformer, KFold, StratifiedKFold, TimeSeriesSplit.
  • Use deepbox/ndarray for tensor.
  • Nine short parts on preprocessing: imputation, feature selection, text vectorizers, polynomial and spline features, median-based and power scalers, and cross-validation splitters. PowerTransformer has standardize set to false by default, while scikit-learn uses true. The example shows both settings.

Source

index.ts
/** * Example 38: Feature Engineering & Preprocessing * * Imputation (SimpleImputer, KNNImputer), feature selection (SelectKBest, * VarianceThreshold), text vectorizers (Count, TF-IDF), PolynomialFeatures, * SplineTransformer, RobustScaler, PowerTransformer and cross-validation splitters. */ import { makeClassification } from "deepbox/datasets";import { tensor } from "deepbox/ndarray";import {  CountVectorizer,  fClassif,  KFold,  KNNImputer,  PolynomialFeatures,  PowerTransformer,  RobustScaler,  SelectKBest,  SimpleImputer,  SplineTransformer,  StratifiedKFold,  TfidfVectorizer,  TimeSeriesSplit,  VarianceThreshold,} from "deepbox/preprocess"; console.log("=".repeat(60));console.log("Example 38: Feature Engineering & Preprocessing");console.log("=".repeat(60)); // ============================================================================// Part 1: SimpleImputer (fill missing values)// ============================================================================console.log("\nPart 1: SimpleImputer");console.log("-".repeat(60)); // SimpleImputer replaces missing values (NaN) using a column statistic or a constantconst XMissing = tensor([  [1, 2, NaN],  [3, NaN, 6],  [7, 8, 9],  [NaN, 5, 3],  [4, 6, 7],]); console.log("Data with missing values:");console.log(XMissing.toString()); // Strategy: mean (default)const impMean = new SimpleImputer({ strategy: "mean" });const XFilledMean = impMean.fitTransform(XMissing);console.log("\nSimpleImputer (mean):");console.log(XFilledMean.toString()); // Strategy: medianconst impMedian = new SimpleImputer({ strategy: "median" });const XFilledMedian = impMedian.fitTransform(XMissing);console.log("SimpleImputer (median):");console.log(XFilledMedian.toString()); // Strategy: constantconst impConst = new SimpleImputer({ strategy: "constant", fillValue: -1 });const XFilledConst = impConst.fitTransform(XMissing);console.log("SimpleImputer (constant=-1):");console.log(XFilledConst.toString()); // ============================================================================// Part 2: KNNImputer (k-nearest neighbors imputation)// ============================================================================console.log("\nPart 2: KNNImputer");console.log("-".repeat(60)); // KNNImputer fills a missing value with the mean of that feature over the k nearest rowsconst knnImp = new KNNImputer({ nNeighbors: 2 });const XFilledKNN = knnImp.fitTransform(XMissing);console.log("KNNImputer (k=2):");console.log(XFilledKNN.toString()); // ============================================================================// Part 3: VarianceThreshold (remove low-variance features)// ============================================================================console.log("\nPart 3: VarianceThreshold");console.log("-".repeat(60)); // VarianceThreshold drops features whose variance is below the thresholdconst XVar = tensor([  [0, 2, 0.1],  [0, 4, 0.2],  [0, 6, 0.1],  [0, 8, 0.3],  [0, 10, 0.2],]); console.log("Original data (feature 0 has zero variance):");console.log(XVar.toString()); const varThresh = new VarianceThreshold({ threshold: 0.01 });const XVarReduced = varThresh.fitTransform(XVar);console.log("\nAfter VarianceThreshold (threshold=0.01):");console.log(XVarReduced.toString());console.log(`  Removed ${XVar.shape[1]! - XVarReduced.shape[1]!} low-variance feature(s)`); // ============================================================================// Part 4: SelectKBest (univariate feature selection)// ============================================================================console.log("\nPart 4: SelectKBest");console.log("-".repeat(60)); // SelectKBest keeps the k features with the highest score. fClassif is the ANOVA F-value.const [XFeat, yFeat] = makeClassification({  nSamples: 100,  nFeatures: 10,  nInformative: 3,  nClasses: 2,  randomState: 42,}); console.log(`Original features: ${XFeat.shape[1]}`); const selector = new SelectKBest({ scoreFunc: fClassif, k: 5 });const XSelected = selector.fitTransform(XFeat, yFeat);console.log(`After SelectKBest (k=5): ${XSelected.shape[1]} features`);console.log(`  Selected feature scores: ${selector.scores.toString()}`); // ============================================================================// Part 5: Text Vectorization// ============================================================================console.log("\nPart 5: Text Vectorization");console.log("-".repeat(60)); // CountVectorizer turns documents into word-count vectorsconst documents = ["the cat sat on the mat", "the dog sat on the log", "cats and dogs are friends"]; console.log("Documents:");for (const doc of documents) {  console.log(`  "${doc}"`);} const countVec = new CountVectorizer();const XCount = countVec.fitTransformText(documents);console.log(`\nCountVectorizer output shape: [${XCount.shape.join(", ")}]`);console.log(`  Vocabulary size: ${countVec.vocabulary.size}`); // TfidfVectorizer weights each count by how rare the word is across documentsconst tfidfVec = new TfidfVectorizer();const XTfidf = tfidfVec.fitTransformText(documents);console.log(`\nTfidfVectorizer output shape: [${XTfidf.shape.join(", ")}]`);console.log("  Words that appear in few documents get higher weights"); // ============================================================================// Part 6: PolynomialFeatures// ============================================================================console.log("\nPart 6: PolynomialFeatures");console.log("-".repeat(60)); // PolynomialFeatures adds powers and products of the input featuresconst XPoly = tensor([  [1, 2],  [3, 4],  [5, 6],]); console.log("Original features:");console.log(XPoly.toString()); const poly = new PolynomialFeatures({ degree: 2, includeBias: false });const XPolyTransformed = poly.fitTransform(XPoly);console.log("\nPolynomialFeatures (degree=2, no bias):");console.log(XPolyTransformed.toString());console.log(  `  Original: ${XPoly.shape[1]} features, expanded: ${XPolyTransformed.shape[1]} features`); // ============================================================================// Part 7: SplineTransformer// ============================================================================console.log("\nPart 7: SplineTransformer");console.log("-".repeat(60)); // SplineTransformer expands one feature into B-spline basis functionsconst XSpline = tensor([[0], [1], [2], [3], [4], [5], [6], [7], [8], [9]]); console.log("Original 1D feature:");console.log(XSpline.toString()); const spline = new SplineTransformer({ nKnots: 4, degree: 3 });const XSplineTransformed = spline.fitTransform(XSpline);console.log("\nSplineTransformer (4 knots, degree 3):");console.log(`  Output shape: [${XSplineTransformed.shape.join(", ")}]`);console.log(`  1 feature becomes ${XSplineTransformed.shape[1]} spline basis functions`); // ============================================================================// Part 8: RobustScaler & PowerTransformer// ============================================================================console.log("\nPart 8: Advanced Scalers");console.log("-".repeat(60)); const XSkewed = tensor([  [1, 100],  [2, 200],  [3, 300],  [100, 400], // outlier in the first feature  [4, 500],]); // RobustScaler centers on the median and scales by the interquartile range, so one outlier barely changes itconst robust = new RobustScaler();const XRobust = robust.fitTransform(XSkewed);console.log("RobustScaler (median and IQR):");console.log(XRobust.toString()); // PowerTransformer applies a power transform that makes each feature more Gaussian.// standardize defaults to false in Deepbox. scikit-learn defaults to true, so pass// { standardize: true } to get zero mean and unit variance as scikit-learn does.const XPositive = tensor([  [1, 10],  [2, 20],  [3, 30],  [4, 40],  [5, 50],]); const power = new PowerTransformer({ method: "yeo-johnson" });const XPower = power.fitTransform(XPositive);console.log("\nPowerTransformer (Yeo-Johnson, standardize=false):");console.log(XPower.toString()); const powerStd = new PowerTransformer({ method: "yeo-johnson", standardize: true });console.log("PowerTransformer (Yeo-Johnson, standardize=true):");console.log(powerStd.fitTransform(XPositive).toString()); // ============================================================================// Part 9: Cross-Validation Splitters// ============================================================================console.log("\nPart 9: Cross-Validation Splitters");console.log("-".repeat(60)); const [XSplit, ySplit] = makeClassification({  nSamples: 20,  nFeatures: 2,  nInformative: 2,  nRedundant: 0,  nClasses: 2,  randomState: 42,}); // KFold: k folds, each used once as the test setconst kfold = new KFold({ nSplits: 5, shuffle: true, randomState: 42 });console.log("KFold (5 splits, shuffled):");let foldNum = 1;for (const { trainIndex, testIndex } of kfold.split(XSplit)) {  console.log(    `  Fold ${foldNum}: train=${trainIndex.length} samples, test=${testIndex.length} samples`  );  foldNum++;} // StratifiedKFold: keeps the class proportions in each foldconst stratKfold = new StratifiedKFold({ nSplits: 5 });console.log("\nStratifiedKFold (5 splits, preserves class balance):");foldNum = 1;for (const { trainIndex, testIndex } of stratKfold.split(XSplit, ySplit)) {  console.log(`  Fold ${foldNum}: train=${trainIndex.length}, test=${testIndex.length}`);  foldNum++;} // TimeSeriesSplit: each test fold comes after its training rows in timeconst tsSplit = new TimeSeriesSplit({ nSplits: 4 });console.log("\nTimeSeriesSplit (4 splits, expanding window):");foldNum = 1;for (const { trainIndex, testIndex } of tsSplit.split(XSplit)) {  console.log(`  Fold ${foldNum}: train=${trainIndex.length}, test=${testIndex.length}`);  foldNum++;} // ============================================================================// Summary// ============================================================================console.log("\nKey Takeaways");console.log("-".repeat(60));console.log("• SimpleImputer: fill NaN with the mean, median, most_frequent value or a constant");console.log("• KNNImputer: fill NaN from the nearest rows");console.log("• VarianceThreshold: drop near-constant features");console.log("• SelectKBest: keep the k features with the best test score");console.log("• CountVectorizer: word counts per document");console.log("• TfidfVectorizer: word counts weighted by rarity");console.log("• SplineTransformer: nonlinear feature expansion with B-splines");console.log("• RobustScaler: scaling that outliers barely affect");console.log("• PowerTransformer: more Gaussian features (standardize is false by default)");console.log("• KFold, StratifiedKFold, TimeSeriesSplit: splitters for different kinds of data"); console.log("\nFeature Engineering Example Complete!");console.log("=".repeat(60));

Output

Console output
Console output only: each transformer's input and result, and the train and test sizes of every fold.
The older name `f_classif` still works. Use `fClassif`.