Feature engineering often involves computationally expensive transformations across large datasets. This example demonstrates parallelizing feature calculation across data segments to dramatically speed up preprocessing pipelines.
Use Case: Machine learning preprocessing, ETL pipelines, data enrichment, time-series feature extraction
Computational Pattern: Data parallelism with chunked processing and aggregation
You have a dataset of 100,000 customer transactions and need to compute multiple engineered features: - Rolling statistics (moving averages, standard deviations) - Lag features (previous N values) - Time-based features (day of week, month, seasonality) - Aggregated customer metrics (lifetime value, purchase frequency) - Text-based features (sentiment scores, keyword extraction) - Categorical encodings (target encoding, frequency encoding)
Computing these features sequentially is slow. Parallelizing by customer ID allows independent processing.
Create synthetic transaction data:
set.seed(123)
# Generate 100,000 transactions for 5,000 customers
n_customers <- 5000
n_transactions <- 100000
transactions <- data.frame(
customer_id = sample(1:n_customers, n_transactions, replace = TRUE),
transaction_date = as.Date("2023-01-01") + sample(0:364, n_transactions, replace = TRUE),
amount = exp(rnorm(n_transactions, log(50), 0.8)), # Log-normal distribution
category = sample(c("groceries", "electronics", "clothing", "restaurants", "other"),
n_transactions, replace = TRUE, prob = c(0.4, 0.1, 0.2, 0.2, 0.1)),
payment_method = sample(c("credit", "debit", "cash"), n_transactions, replace = TRUE),
is_online = sample(c(TRUE, FALSE), n_transactions, replace = TRUE, prob = c(0.3, 0.7))
)
# Sort by customer and date
transactions <- transactions[order(transactions$customer_id, transactions$transaction_date), ]
cat(sprintf("Dataset: %s transactions for %s customers\n",
format(nrow(transactions), big.mark = ","),
format(length(unique(transactions$customer_id)), big.mark = ",")))
cat(sprintf("Date range: %s to %s\n",
min(transactions$transaction_date),
max(transactions$transaction_date)))Output:
Dataset: 100,000 transactions for 5,000 customers
Date range: 2023-01-01 to 2023-12-31Define a function that computes features for a set of customers:
engineer_features <- function(customer_ids, transactions_data) {
# Filter to specified customers
customer_data <- transactions_data[transactions_data$customer_id %in% customer_ids, ]
if (nrow(customer_data) == 0) {
return(NULL)
}
# Sort by customer and date
customer_data <- customer_data[order(customer_data$customer_id,
customer_data$transaction_date), ]
# Compute features for each customer
features_list <- lapply(customer_ids, function(cid) {
cust_txn <- customer_data[customer_data$customer_id == cid, ]
if (nrow(cust_txn) == 0) {
return(NULL)
}
# Basic statistics
total_transactions <- nrow(cust_txn)
total_spend <- sum(cust_txn$amount)
avg_transaction <- mean(cust_txn$amount)
# Time-based features
first_purchase <- min(cust_txn$transaction_date)
last_purchase <- max(cust_txn$transaction_date)
days_active <- as.numeric(difftime(last_purchase, first_purchase, units = "days"))
purchase_frequency <- if (days_active > 0) total_transactions / days_active else 0
# Rolling statistics (last 30 days)
recent_cutoff <- last_purchase - 30
recent_txn <- cust_txn[cust_txn$transaction_date > recent_cutoff, ]
recent_spend <- if (nrow(recent_txn) > 0) sum(recent_txn$amount) else 0
recent_count <- nrow(recent_txn)
# Category preferences
category_counts <- table(cust_txn$category)
favorite_category <- names(category_counts)[which.max(category_counts)]
category_diversity <- length(unique(cust_txn$category))
# Payment behavior
online_pct <- mean(cust_txn$is_online) * 100
credit_pct <- mean(cust_txn$payment_method == "credit") * 100
# Volatility metrics
amount_sd <- sd(cust_txn$amount)
amount_cv <- amount_sd / avg_transaction # Coefficient of variation
# Time patterns
weekday <- as.POSIXlt(cust_txn$transaction_date)$wday
weekend_pct <- mean(weekday %in% c(0, 6)) * 100
# RFM features (Recency, Frequency, Monetary)
recency <- as.numeric(difftime(as.Date("2024-01-01"), last_purchase, units = "days"))
data.frame(
customer_id = cid,
total_transactions = total_transactions,
total_spend = total_spend,
avg_transaction = avg_transaction,
days_active = days_active,
purchase_frequency = purchase_frequency,
recent_spend_30d = recent_spend,
recent_count_30d = recent_count,
favorite_category = favorite_category,
category_diversity = category_diversity,
online_pct = online_pct,
credit_pct = credit_pct,
amount_sd = amount_sd,
amount_cv = amount_cv,
weekend_pct = weekend_pct,
recency_days = recency,
stringsAsFactors = FALSE
)
})
# Combine results
do.call(rbind, features_list[!sapply(features_list, is.null)])
}Test feature engineering locally on a subset:
# Process 500 customers locally
set.seed(456)
sample_customers <- sample(unique(transactions$customer_id), 500)
cat(sprintf("Processing features for %d customers locally...\n", length(sample_customers)))
local_start <- Sys.time()
local_features <- engineer_features(sample_customers, transactions)
local_time <- as.numeric(difftime(Sys.time(), local_start, units = "secs"))
cat(sprintf("[OK] Completed in %.2f seconds\n", local_time))
cat(sprintf(" Generated %d features per customer\n", ncol(local_features) - 1))
cat(sprintf(" Estimated time for all %s customers: %.1f minutes\n",
format(n_customers, big.mark = ","),
local_time * (n_customers / 500) / 60))Illustrative output:
Processing features for 500 customers locally...
[OK] Completed in 3.8 seconds
Generated 15 features per customer
Estimated time for all 5,000 customers: 38.0 secondsFor all 5,000 customers locally: ~38 seconds
Process all customers in parallel on AWS:
# Split customers into chunks
all_customers <- unique(transactions$customer_id)
n_workers <- 20
cat(sprintf("Processing features for %s customers on %d workers...\n",
format(length(all_customers), big.mark = ","),
n_workers))
# Create chunks of customer IDs
chunk_size <- ceiling(length(all_customers) / n_workers)
customer_chunks <- split(all_customers,
ceiling(seq_along(all_customers) / chunk_size))
cloud_start <- Sys.time()
# Process chunks in parallel
results <- starburst_map(
customer_chunks,
engineer_features,
transactions_data = transactions,
workers = n_workers,
cpu = 2,
memory = "4GB"
)
cloud_time <- as.numeric(difftime(Sys.time(), cloud_start, units = "secs"))
cat(sprintf("\n[OK] Completed in %.1f seconds\n", cloud_time))
# Combine results
features <- do.call(rbind, results)Illustrative output:
[Starting] Starting starburst cluster with 20 workers
[Status] Processing 20 items with 20 workers
[Starting] Submitting 20 tasks...
[Wait] Progress: 20/20 (6.2s)
[OK] Completed in 6.2 seconds
[Cost] Estimated cost: $0.003Examine the engineered features:
cat("\n=== Feature Engineering Results ===\n\n")
cat(sprintf("Total customers processed: %s\n",
format(nrow(features), big.mark = ",")))
cat(sprintf("Features generated per customer: %d\n", ncol(features) - 1))
cat(sprintf("\nFeature names:\n"))
print(names(features)[names(features) != "customer_id"])
# Summary statistics
cat("\n=== Feature Summary Statistics ===\n\n")
cat(sprintf("Average transactions per customer: %.1f\n",
mean(features$total_transactions)))
cat(sprintf("Average total spend: $%.2f\n",
mean(features$total_spend)))
cat(sprintf("Average transaction value: $%.2f\n",
mean(features$avg_transaction)))
cat(sprintf("\nMost common category: %s\n",
names(sort(table(features$favorite_category), decreasing = TRUE)[1])))
cat(sprintf("Average category diversity: %.1f categories\n",
mean(features$category_diversity)))
cat(sprintf("\nAverage online purchase rate: %.1f%%\n",
mean(features$online_pct)))
cat(sprintf("Average credit card usage: %.1f%%\n",
mean(features$credit_pct)))
cat(sprintf("\nMean purchase frequency: %.3f transactions/day\n",
mean(features$purchase_frequency, na.rm = TRUE)))
cat(sprintf("Mean recency: %.0f days since last purchase\n",
mean(features$recency_days)))
# Identify high-value segments
high_value <- features[features$total_spend > quantile(features$total_spend, 0.9), ]
cat(sprintf("\n=== Top 10%% Customers (by spend) ===\n"))
cat(sprintf("Count: %d customers\n", nrow(high_value)))
cat(sprintf("Average total spend: $%.2f\n", mean(high_value$total_spend)))
cat(sprintf("Average transactions: %.1f\n", mean(high_value$total_transactions)))
cat(sprintf("Average online rate: %.1f%%\n", mean(high_value$online_pct)))Illustrative output:
=== Feature Engineering Results ===
Total customers processed: 5,000
Features generated per customer: 15
Feature names:
[1] "total_transactions" "total_spend" "avg_transaction"
[4] "days_active" "purchase_frequency" "recent_spend_30d"
[7] "recent_count_30d" "favorite_category" "category_diversity"
[10] "online_pct" "credit_pct" "amount_sd"
[13] "amount_cv" "weekend_pct" "recency_days"
=== Feature Summary Statistics ===
Average transactions per customer: 20.0
Average total spend: $1,002.45
Average transaction value: $50.12
Most common category: groceries
Average category diversity: 3.8 categories
Average online purchase rate: 30.2%
Average credit card usage: 33.1%
Mean purchase frequency: 0.055 transactions/day
Mean recency: 183 days since last purchase
=== Top 10% Customers (by spend) ===
Count: 500 customers
Average total spend: $2,845.23
Average transactions: 51.2
Average online rate: 28.9%| Method | Customers | Compute time | Cost |
|---|---|---|---|
| Local | 500 | 3.8 sec | $0 |
| Local (est.) | 5,000 | 38 sec | $0 |
| staRburst | 5,000 | 6.2 sec (compute only) | $0.003 |
Reading these numbers honestly: the staRburst row is compute time only on a warm pool and excludes the one-time startup (~2 minutes) and image pull. For a job that finishes in ~38 seconds locally, that startup makes local the faster choice — these figures are illustrative, not a net-win claim. The payoff comes at scale (100K+ customers or heavier per-record work). See
vignette("performance").
Key Insights: - Feature engineering parallelizes near-linearly once workers are warm - Minimal per-task cost - The win is at scale, where startup overhead is amortized away - Can process 100 customers/second with 20 workers
Extend the feature function for machine learning:
engineer_ml_features <- function(customer_ids, transactions_data) {
base_features <- engineer_features(customer_ids, transactions_data)
if (is.null(base_features)) {
return(NULL)
}
# Add derived features for ML
base_features$clv_estimate <- base_features$avg_transaction *
base_features$purchase_frequency * 365
base_features$engagement_score <- scale(base_features$total_transactions) +
scale(base_features$category_diversity)
base_features$recency_score <- -scale(base_features$recency_days)
base_features$is_high_value <- base_features$total_spend >
median(base_features$total_spend)
base_features$purchase_momentum <- base_features$recent_count_30d /
(base_features$total_transactions + 1)
base_features
}
# Process with enhanced features
ml_results <- starburst_map(
customer_chunks,
engineer_ml_features,
transactions_data = transactions,
workers = 20,
cpu = 2,
memory = "4GB"
)
ml_features <- do.call(rbind, ml_results)Good fit: - Large datasets requiring expensive transformations - Independent processing by group (customer, product, region) - Multiple feature types needing calculation - Iterative feature engineering pipelines - Real-time feature stores with batch updates
Not ideal: - Simple vectorized operations (use dplyr/data.table) - Features requiring global statistics - Very small datasets (< 1,000 groups) - Real-time single-record scoring
The complete runnable script is available at:
Run it with:
Related examples: - Grid Search - Hyperparameter tuning with features - Risk Modeling - Advanced customer analytics - Bootstrap CI - Feature importance estimation