How to Split a Dataset: Train, Validation & Test Sets Explained | Day 22

Day 22
ML Fundamentals

Kal humne samjha tha ki Dataset kya hota hai. Lekin jab ek clean dataset taiyar ho jata hai, toh kya hum poora data ek saath Machine Learning model ko sikhane ke liye de dete hain? Jawab hai: Bilkul nahi!


Namaste! Main hoon Coach Krish.

Imagine kijiye ki aap ek student ko Final Exam ki taiyari karwa rahe hain. Agar aap use wahi exact questions practice paper mein de denge jo kal exam mein aane wale hain, toh student test mein 100 out of 100 le aayega. Lekin kya iska matlab ye hai ki student ko subject genuinely samajh aa gaya hai? Nahi, usne bas answers ko ratta (memorize) mar liya hai.

Machine Learning models ke saath bhi exact yahi problem hoti hai. Isiliye hum Dataset Splitting ka use karte hain. Aaj ke lesson mein hum sikhenge ki dataset ko Train, Validation, aur Test sets mein kaise aur kyun split kiya jata hai.

Dataset Split Karna Kyun Zaroori Hai?

Machine Learning ka ultimate goal sirf purane data par achha perform karna nahi hai, balki unseen data (naye data) par accurate predictions dena hai. Is ability ko hum ML mein Generalization kehte hain.

Agar hum poora data model ko train karne mein use kar lenge, toh hum kabhi test hi nahi kar paayenge ki model naye cases par kaisa perform karega. Isiliye dataset ko hum alag-alag parts mein divide karte hain.

๐Ÿง  Mental Model: The Exam Preparation Framework

Dataset split ko samajhne ke liye sabse best mental model hai Exam System:

  • Training Set (Textbook & Homework): Is data se model patterns seekhta hai. Model inputs aur correct targets dono ko baar-baar dekhta hai.
  • Validation Set (Practice / Mock Tests): Training ke dauran hum model ka test lene ke liye is set ka use karte hain taaki uske parameters ko tune kiya ja sake (jaise padhai ki strategy adjust karna).
  • Test Set (Final Board Exam): Ye data model se poori tarah chhupa kar rakha jata hai. Model jab fully train ho jata hai, tab final performance score napne ke liye iska ek hi baar evaluation hota hai.

Standard Dataset Split Ratios

Data ko kis proportion mein split karna chahiye, ye aapke dataset ke total size par depend karta hai. Standard practice ke acche guidelines yahan hain:

Dataset Size Train Set Validation Set Test Set
Small to Medium (< 100,000 rows) 70% – 80% 10% – 15% 10% – 15%
Large Data / Deep Learning (> 1M rows) 98% 1% 1%

Note: Modern Deep Learning mein jab hamare paas millions of rows hoti hain, toh 1% data bhi thousands of samples ke barabar hota hai, jo validation ya test ke liye kaafi hota hai.

Technical Example: Python Scikit-Learn Split

Python mein popular library scikit-learn ka use karke dataset ko split karna bahut simple hai. Dekhiye code kaise kaam karta hai:

# Importing required library
from sklearn.model_selection import train_test_split

# Assume X contains Features and y contains Target Labels
# Step 1: Split data into Train (80%) and Temp (20%)
X_train, X_temp, y_train, y_temp = train_test_split(X, y, test_size=0.20, random_state=42)

# Step 2: Split Temp data into Validation (50% of Temp = 10% total) and Test (50% of Temp = 10% total)
X_val, X_test, y_val, y_test = train_test_split(X_temp, y_temp, test_size=0.50, random_state=42)

print(f"Train size: {len(X_train)}, Val size: {len(X_val)}, Test size: {len(X_test)}")

Yahan random_state=42 ka use split ko repeatable aur consistent banane ke liye kiya gaya hai taaki har baar same random split mile.

Common Mistakes: Watch Out for Data Leakage

Beginners aksar ye 2 sabse badi mistakes karte hain:

๐Ÿšจ 1. Data Leakage (Cheating in Training)

Agar Test Set ki information (jaise averages, scaling parameters, ya duplicate rows) Training Set mein kisi tarah mix ho jaye, toh use Data Leakage kehte hain. Isse model ka training score artificially high dikhta hai lekin real-world deployment par model fail ho jata hai.

โš ๏ธ 2. Random Splitting on Time Series Data

Agar aap Stock Prices ya Weather forecast jaise time-dependent data par kaam kar rahe hain, toh random split mat kijiye! Time-series data ko humesha Chronological Split (purana data = Train, naya data = Test) kiya jata hai. Future ka data past mein leak nahi hona chahiye.

๐Ÿ’ฌ Coach Krish’s Coach Note

“Humesha yaad rakhiye: Test set ko tab tak ek lock box mein rakhein jab tak aapka model training aur tuning stage se poori tarah nikal na jaye. Agar aap test set ke results ko dekh kar baar-baar hyperparameter change karenge, toh aapka validation aur test set ka fark khatam ho jayega. Keep Test Data Sacred!”

Quick Recall Checkpoint

Apne understanding ko test karne ke liye niche diye gaye questions ke answers baare mein sochiye:

Q1: Validation Set aur Test Set mein main difference kya hai?

Validation Set ka use model training ke dauran tuning aur evaluation ke liye hota hai. Test Set ka use sirf training khatam hone ke baad final score check karne ke liye hota hai.

Q2: Data Leakage se model par kya effect padta hai?

Data Leakage se model testing par bohot accha dikhta hai (over-optimistic score), lekin actual application par totally fail ho jata hai.

Q3: Time-Series data mein random split kyun nahi karna chahiye?

Kyunki random split se future information past mein leak ho jayegi, jo Real World mein possible nahi hoti. Temporal order maintain karna zaroori hai.

๐Ÿ“Œ Key Takeaways

  • Dataset split ka main purpose model ki Generalization Ability verify karna hota hai.
  • Train Set: Model seekhta hai.
  • Validation Set: Model tune aur select hota hai.
  • Test Set: Final evaluation (unseen real-world exam).
  • Pre-processing (Jaise Feature Scaling) humesha split karne ke BAAD aur sirf Training set par fit honi chahiye.

What Comes Next?

Aaj humne sikha ki data ko Train aur Test mein split kaise kiya jata hai. Lekin tab kya hota hai jab humara model Training Data par toh 99% accuracy deta hai, lekin Test Data par fail ho jata hai? Is critical machine learning phenomenon ko Overfitting kehte hain.

Kal ke Day 23 lesson mein hum samjhenge: Overfitting Explained โ€” Kyun Model Ratta Maar Leta Hai?

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top