recipes-patterns

# Recipes Feature Engineering Patterns ## Overview Comprehensive patterns for feature engineering using the recipes package. Covers preprocessing steps for numeric, categorical, and text data while preventing information leakage. ## Recipe Fundamentals ### Basic Recipe Structure ```r library(recipes) # Initialize recipe with formula rec <- recipe(outcome ~ ., data = training_data) # Or with explicit roles rec <- recipe(training_data) |> update_role(outcome, new_role = "outcome") |> update_role(id_column, new_role = "ID") |> update_role(-outcome, -id_column, new_role = "predictor") ``` ### Selector Functions ```r # Type-based selectors all_predictors() all_outcomes() all_numeric_predictors() all_nominal_predictors() all_numeric() all_nominal() # Name-based selectors starts_with("prefix_") ends_with("_suffix") contains("pattern") matches("regex") one_of(c("var1", "var2")) ``` ## Numeric Preprocessing ### Normalization and Scaling ```r rec <- recipe(outcome ~ ., data = train) |> # Center and scale (z-score) step_normalize(all_numeric_predictors()) |> # Scale to [0, 1] step_range(all_numeric_predictors(), min = 0, max = 1) |> # Center only step_center(all_numeric_predictors()) |> # Scale only step_scale(all_numeric_predictors()) ``` ### Transformations for Normality ```r rec <- recipe(outcome ~ ., data = train) |> # Yeo-Johnson (handles zero and negative values) step_YeoJohnson(all_numeric_predictors()) |> # Box-Cox (positive values only) step_BoxCox(positive_vars) |> # Log transformation step_log(skewed_vars, base = 10) |> # Square root step_sqrt(count_vars) ``` ### Spline and Polynomial Features ```r rec <- recipe(outcome ~ ., data = train) |> # Natural splines step_ns(continuous_var, deg_free = 5) |> # B-splines step_bs(continuous_var, deg_free = 5, degree = 3) |> # Polynomial features step_poly(continuous_var, degree = 3) ``` ## Categorical Encoding ### Dummy Variables ```r rec <- recipe(outcome ~ ., data = train) |> # One-hot encoding (drop first level) step_dummy(all_nominal_predictors()) |> # Keep all levels step_dummy(all_nominal_predictors(), one_hot = TRUE) ``` ### Handling Rare Categories ```r rec <- recipe(outcome ~ ., data = train) |> # Pool infrequent levels step_other(categorical_var, threshold = 0.05, other = "other") |> # Handle novel levels in new data step_novel(all_nominal_predictors()) |> # Convert NA to explicit level step_unknown(all_nominal_predictors()) ``` ### Target Encoding (embed package) ```r library(embed) rec <- recipe(outcome ~ ., data = train) |> # Likelihood encoding step_lencode_glm(high_cardinality_var, outcome = vars(outcome)) |> # Mixed model encoding (for hierarchical data) step_lencode_mixed(category, outcome = vars(outcome)) |> # Weight of evidence step_woe(categorical_var, outcome = vars(binary_outcome)) ``` ## Missing Data Handling ### Imputation Methods ```r rec <- recipe(outcome ~ ., data = train) |> # Simple imputation step_impute_mean(all_numeric_predictors()) |> step_impute_median(all_numeric_predictors()) |> step_impute_mode(all_nominal_predictors()) |> # KNN imputation step_impute_knn(all_predictors(), neighbors = 5) |> # Bagged tree imputation step_impute_bag(all_predictors()) |> # Linear model imputation step_impute_linear(numeric_var, impute_with = imp_vars(predictor1, predictor2)) ``` ### Missing Indicators ```r rec <- recipe(outcome ~ ., data = train) |> # Create indicator for missingness step_indicate_na(all_predictors()) |> # Then impute step_impute_median(all_numeric_predictors()) |> step_impute_mode(all_nominal_predictors()) ``` ## Dimensionality Reduction ### PCA and Related Methods ```r rec <- recipe(outcome ~ ., data = train) |> step_normalize(all_numeric_predictors()) |> # Principal Component Analysis step_pca(all_numeric_predictors(), num_comp = 5) |> # Or keep components explaining variance threshold step_pca(all_numeric_predictors(), threshold = 0.95) ``` ### Other Reduction Methods ```r library(embed) rec <- recipe(outcome ~ ., data = train) |> # UMAP step_umap(all_numeric_predictors(), num_comp = 2) |> # Kernel PCA step_kpca(all_numeric_predictors(), num_comp = 5) ``` ## Interactions and Combinations ```r rec <- recipe(outcome ~ ., data = train) |> # Create interaction terms step_interact(terms = ~ var1:var2) |> # Multiple interactions step_interact(terms = ~ starts_with("x"):starts_with("z")) |> # Ratios step_ratio(numerator = denom_vars(var1), denom = denom_vars(var2)) ``` ## Feature Selection ```r rec <- recipe(outcome ~ ., data = train) |> # Remove zero variance step_zv(all_predictors()) |> # Remove near-zero variance step_nzv(all_predictors(), freq_cut = 95/5, unique_cut = 10) |> # Remove highly correlated step_corr(all_numeric_predictors(), threshold = 0.9) |> # Remove linear combinations step_lincomb(all_numeric_predictors()) ``` ## Class Imbalance (themis) ```r library(themis) rec <- recipe(outcome ~ ., data = train) |> # Downsample majority class step_downsample(outcome) |> # Upsample minority class step_upsample(outcome) |> # SMOTE step_smote(outcome) |> # ADASYN step_adasyn(outcome) ``` ## Date/Time Features ```r rec <- recipe(outcome ~ ., data = train) |> # Extract date components step_date(date_var, features = c("year", "month", "dow", "doy")) |> # Holiday indicators step_holiday(date_var, holidays = c("USChristmasDay", "USNewYearsDay")) |> # Time components step_time(datetime_var, features = c("hour", "minute")) ``` ## Text Features (textrecipes) ```r library(textrecipes) rec <- recipe(outcome ~ ., data = train) |> step_tokenize(text_var) |> step_stopwords(text_var) |> step_stem(text_var) |> step_ngram(text_var, num_tokens = 2) |> step_tfidf(text_var, max_tokens = 100) ``` ## Recipe Execution ```r # Prepare recipe (estimate parameters from training data) prepped_rec <- prep(rec, training = train_data) # Apply to training data train_processed <- bake(prepped_rec, new_data = NULL) # or juice(prepped_rec) # Apply to new data test_processed <- bake(prepped_rec, new_data = test_data) # Inspect recipe tidy(prepped_rec) tidy(prepped_rec, number = 1) # specific step ``` ## Step Ordering Best Practices ```r recipe(outcome ~ ., data = train) |> # 1. Handle roles and IDs update_role(id, new_role = "ID") |> # 2. Impute missing values first step_impute_median(all_numeric_predictors()) |> step_impute_mode(all_nominal_predictors()) |> # 3. Handle individual variable issues step_other(all_nominal_predictors(), threshold = 0.05) |> step_novel(all_nominal_predictors()) |> # 4. Transform numeric variables step_YeoJohnson(all_numeric_predictors()) |> # 5. Create interactions before encoding step_interact(terms = ~ var1:var2) |> # 6. Encode categorical variables step_dummy(all_nominal_predictors()) |> # 7. Normalize (after dummy coding) step_normalize(all_numeric_predictors()) |> # 8. Feature selection last step_zv(all_predictors()) |> step_corr(all_numeric_predictors()) ``` ## Key Principles 1. **Prevent leakage**: All statistics computed from training data only 2. **Order matters**: Impute → Transform → Encode → Normalize → Select 3. **Use selectors**: More maintainable than listing variable names 4. **Document decisions**: Comment why each step is included 5. **Test on holdout**: Verify recipe generalizes to new data