Train, validation, and test splits

Split data so you train parameters, tune on validation, and report final performance on a held-out test set touched once. Random splits fail when data has time or group structure.

Three sets

• Train — fit model weights
• Validation — pick hyperparameters, early stopping
• Test — unbiased estimate before launch (use sparingly)

Split strategies

• Random — IID rows (rare in production)
• Time-based — train on past, validate on future
• Group — all rows from one user in one split only

Split pseudocode

# 70/15/15 split concept
n = 1000
train_end = int(n * 0.70)
val_end = int(n * 0.85)
print("train:", train_end, "val:", val_end - train_end, "test:", n - val_end)

Practice: Optional snippets use pandas-style pseudocode—run with Pandas locally if you want tactile practice.

Important interview questions and answers

1. Q: Why not tune on test?
   A: Test becomes validation—optimistic bias on final metrics.
2. Q: Time split when?
   A: User behavior drifts; future must not appear in training features.

Self-check

1. What is each split used for?
2. When use group split instead of random?

Tip: Use time-based splits when user behavior drifts seasonally.