Introduction — Problem Kya Hai?

Socho tumhare paas California ke hazaron ghar ka data hai — unki location, age, rooms, income of people — aur tumhara kaam hai yeh predict karna ki ek naye ghar ki median price kya hogi?

Yeh ek classic Regression problem hai. Classification mein hum categories predict karte hain (spam/not spam), lekin regression mein hum ek continuous number predict karte hain — jaise ghar ki keemat.

🎯 Goal

California Housing dataset ka use karke ek Random Forest Regressor model train karna jo diye gaye features ke basis par median_house_value predict kare.

Is blog mein hum do approaches cover karenge — pehle manual preprocessing, phir usi kaam ko Sklearn Pipeline ke saath professionally karna. End mein clearly samjhayenge ki dono mein kya difference hai aur kaunsa better hai.

Dataset ka Overview

Hum California Housing dataset use kar rahe hain. Yeh dataset real-world California census data par based hai. Isme 20,640 rows aur 10 columns hain.

File:- housing.csv

Columns ki List:

num_columns = ['longitude', 'latitude', 'housing_median_age', 'total_rooms', 'total_bedrooms', 'population', 'households', 'median_income']
cat_columns = ['ocean_proximity']

Categorical Feature (Text)

ocean_proximity

Values: NEAR BAY, INLAND, <1H OCEAN, NEAR OCEAN

🎯 Target Variable (Jo Predict Karna Hai)

median_house_value

⚠️

Important: total_bedrooms column mein 207 null values hain. Inhe handle karna zaroori hai warna model error dega!

Part 1 — Manual Preprocessing + Model Training

Pehle hum step-by-step manually data prepare karenge. Yeh approach beginners ke liye samajhna aasaan hai kyunki har step clearly alag hai.

Data Load Karo

Pandas se CSV file load karo aur pehli 2 rows dekho ki data kaisa hai.

Python

import pandas as pd

data = pd.read_csv('housing.csv')
data.head(2)

Categorical Column Encode Karo — OneHotEncoding

Machine Learning models numbers samajhte hain, text nahi. Isliye ocean_proximity (text column) ko numbers mein convert karna padta hai — yeh kaam OneHotEncoder karta hai.

💡 OneHotEncoding Kya Hoti Hai?

Agar ocean_proximity mein 4 categories hain (NEAR BAY, INLAND, etc.), toh OneHotEncoding 4 naye columns banata hai — har category ke liye ek. Agar ghar us category ka hai toh 1, warna 0.

Example: "NEAR BAY" → [1, 0, 0, 0] | "INLAND" → [0, 1, 0, 0]

Python

# Categorical columns define karo
cat_columns = ['ocean_proximity']

from sklearn.preprocessing import OneHotEncoder

encoder = OneHotEncoder()
encoded_cat = encoder.fit_transform(data[cat_columns])

# Numpy array ko DataFrame mein convert karo
encoded_cat_df = pd.DataFrame(
    encoded_cat.toarray(),
    columns=encoder.get_feature_names_out(cat_columns)
)

# Original column drop karo, encoded columns add karo
data = pd.concat([data.drop(columns=cat_columns), encoded_cat_df], axis=1)

Null Values Handle Karo

total_bedrooms mein missing values hain. Hum unhe median se fill karte hain — yeh ek safe aur common approach hai.

💡 Median vs Mean — Kaunsa Better?

Mean (average) outliers se affect hota hai. Agar kisi block mein 10,000 bedrooms hain (outlier), toh mean bahut badh jayega.

Median middle value hota hai — outliers se affect nahi hota. Housing data mein outliers common hain, isliye median preferred hai.

Python

# Null values ko median se fill karo
data['total_bedrooms'] = data['total_bedrooms'].fillna(
    data['total_bedrooms'].median()
)

Numerical Features Scale Karo — StandardScaler

Kuch features bahut bade numbers mein hain (population: 10,000+) aur kuch chote mein (median_income: 1–15). Yeh imbalance model ko confuse kar sakta hai. StandardScaler sab values ko same scale par le aata hai.

💡 StandardScaler Kya Karta Hai?

Har feature ka mean 0 aur standard deviation 1 kar deta hai.
Formula: z = (x - mean) / std_dev

Iska faida: distance-based algorithms (aur kuch tree-based bhi) better kaam karte hain jab sab features same scale par hoon.

Python

from sklearn.preprocessing import StandardScaler

num_columns = ['longitude', 'latitude', 'housing_median_age',
               'total_rooms', 'total_bedrooms', 'population',
               'households', 'median_income']

scaler = StandardScaler()
data[num_columns] = scaler.fit_transform(data[num_columns])

Train-Test Split Karo

Data ko 80% training aur 20% testing mein baanto. Model training data se seekhega aur testing data par evaluate hoga.

Python

from sklearn.model_selection import train_test_split

X = data.drop(columns=['median_house_value'])  # Features
y = data['median_house_value']                    # Target

X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.2, random_state=42
)

Random Forest Model Train Karo

Random Forest ek powerful algorithm hai jo kai decision trees banata hai aur unka average nikalta hai. Yeh overfitting se bachata hai aur accuracy achchi deta hai.

🌲 Random Forest — Simple Explanation

Ek akela decision tree kabhi kabhi galat decision leta hai — jaise ek judge.

Random Forest mein 100 trees (n_estimators=100) hain — 100 judges. Har tree thoda alag data aur features dekhta hai. Final prediction sabka average hota hai — democratic voting jaisa!

Real world analogy: Ek doctor ki jagah 100 doctors ki second opinion lena.

Python

from sklearn.ensemble import RandomForestRegressor

model = RandomForestRegressor(n_estimators=100, random_state=42)
model.fit(X_train, y_train)

Model Evaluation — MAE, MSE, R² Score

Model train hone ke baad sabse important question hai — "Yeh kitna accurate hai?" Regression ke liye hum teen metrics use karte hain:

MAE

Mean Absolute Error

Average absolute difference predicted aur actual values ke beech. Lower = better. Easy to understand!

MSE

Mean Squared Error

Errors ko square karke average. Bade errors ko zyada punish karta hai. Lower = better.

R² Score

0 to 1

Model ne variance ka kitna hissa explain kiya. 1 = perfect, 0 = model useless. Higher = better.

⚠️ Common Mistake — Accuracy Use Karna

Bahut beginners regression problem mein accuracy() use karne ki koshish karte hain — yeh bilkul galat hai!

Accuracy classification ke liye hai (spam/not spam). Regression mein hum exact number predict karte hain — toh MAE, MSE, R² use karo.

Python

from sklearn.metrics import mean_absolute_error, mean_squared_error, r2_score

y_pred = model.predict(X_test)

mae = mean_absolute_error(y_test, y_pred)
mse = mean_squared_error(y_test, y_pred)
r2  = r2_score(y_test, y_pred)

print(f'MAE: {mae:.2f}')
print(f'MSE: {mse:.2f}')
print(f'R² Score: {r2:.4f}')

🧠 R² Score Interpret Kaise Karo?

R² = 0.80 ka matlab: model ne target variable ka 80% variance explain kiya.
R² = 0.50 — average model. R² = 0.85+ — good model for housing data.
R² = 1.0 — perfect (suspicious — overfitting check karo!)
R² negative — model se bhi bekar simple average hota.

New Data Point par Prediction

Ek naye ghar ka data diya gaya hai jiske baare mein model ne pehle kabhi nahi dekha — ab hum predict karenge ki uski price kya hogi.

Lekin problem yeh hai — new data raw hai. Hume wahi saari preprocessing manually apply karni padegi jo training mein ki thi.

Python

# New data point define karo
new_data = {
    'longitude': -122.23,  'latitude': 37.88,
    'housing_median_age': 41,
    'total_rooms': 880,     'total_bedrooms': 129,
    'population': 322,     'households': 126,
    'median_income': 8.3252,
    'ocean_proximity': 'NEAR BAY'
}

new_data_df = pd.DataFrame([new_data])

# Step 1: Categorical encode karo (same encoder use karo!)
encoded_new_cat = encoder.transform(new_data_df[cat_columns])
encoded_new_cat_df = pd.DataFrame(
    encoded_new_cat.toarray(),
    columns=encoder.get_feature_names_out(cat_columns)
)

# Step 2: Drop original, concat encoded
new_data_df = pd.concat(
    [new_data_df.drop(columns=cat_columns), encoded_new_cat_df], axis=1
)

# Step 3: Scale (same scaler use karo!)
new_data_df[num_columns] = scaler.transform(new_data_df[num_columns])

# Step 4: Predict!
predicted_value = model.predict(new_data_df)
print(f'Predicted Price: ₹{predicted_value[0]:,.0f}')

Manual Approach ka Bada Problem

⚠️ Yahan ek serious problem hai!

Dekho — new data predict karne ke liye humne manually 4 alag steps kiye:
encoder.transform → concat → scaler.transform → model.predict

Agar tum production mein yeh code deploy karo aur koi ek step bhool jao — model galat results dega bina koi error bataye. Yeh debugging nightmare hai!

Manual Approach ki Problems

Human error: Koi step miss ho sakta hai (especially scaling ya encoding)
Order matter karta hai: Agar scale pehle karo aur encode baad mein — wrong results
Objects separately store karne padte hain: encoder, scaler, model — teeno alag save/load karne padte hain
Team mein share karna mushkil: Dusra developer sab steps nahi jaanta
Repetitive code: Training aur prediction dono jagah same steps likhne padte hain

Solution? → Sklearn Pipeline! 🚀

Part 2 — Sklearn Pipeline (Professional Way)

Pipeline ek aisa container hai jo saari preprocessing steps aur model ko ek single object mein pack kar deta hai. Ek baar fit karo, ek baar predict karo — baaki sab automatically hota hai.

Step-by-Step Pipeline Code:

Python — Full Pipeline

import pandas as pd
from sklearn.pipeline import Pipeline
from sklearn.compose import ColumnTransformer
from sklearn.impute import SimpleImputer
from sklearn.preprocessing import StandardScaler, OneHotEncoder
from sklearn.ensemble import RandomForestRegressor
from sklearn.model_selection import train_test_split

data = pd.read_csv('housing.csv')

cat_columns = ['ocean_proximity']
num_columns = ['longitude', 'latitude', 'housing_median_age',
               'total_rooms', 'total_bedrooms', 'population',
               'households', 'median_income']

# Numerical pipeline: null fill → scale
num_transformer = Pipeline(steps=[
    ('imputer', SimpleImputer(strategy='median')),
    ('scaler', StandardScaler())
])

# Categorical pipeline: encode
cat_transformer = Pipeline(steps=[
    ('encoder', OneHotEncoder(handle_unknown='ignore'))
])

# Dono transformers combine karo
preprocessor = ColumnTransformer(transformers=[
    ('num', num_transformer, num_columns),
    ('cat', cat_transformer, cat_columns)
])

# Final pipeline: preprocessor + model
pipeline = Pipeline(steps=[
    ('preprocessor', preprocessor),
    ('model', RandomForestRegressor(n_estimators=100, random_state=42))
])

# Train
X = data.drop(columns=['median_house_value'])
y = data['median_house_value']
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)

pipeline.fit(X_train, y_train)

Evaluation — Same Metrics:

Python

from sklearn.metrics import mean_absolute_error, mean_squared_error, r2_score

y_pred = pipeline.predict(X_test)

print(f"MAE:      {mean_absolute_error(y_test, y_pred):.2f}")
print(f"MSE:      {mean_squared_error(y_test, y_pred):.2f}")
print(f"R² Score: {r2_score(y_test, y_pred):.4f}")

New Data Predict Karna — Kitna Aasaan!

Pipeline ke saath new data prediction sirf 3 lines ka kaam hai:

Python

new_data = {
    'longitude': -122.23,  'latitude': 37.88,
    'housing_median_age': 41,
    'total_rooms': 880,     'total_bedrooms': 129,
    'population': 322,     'households': 126,
    'median_income': 8.3252,
    'ocean_proximity': 'NEAR BAY'
}

new_data_df = pd.DataFrame([new_data])

# BAS! Pipeline sab kuch handle kar lega automatically ✨
predicted_value = pipeline.predict(new_data_df)
print(f'Predicted Price: {predicted_value[0]:,.0f}')

Pipeline ka Jadoo

Raw data diya → pipeline ne khud null fill kiya → encoding ki → scaling ki → model se predict karaya.

No manual steps. No human error. Clean code.

Pro Tips & Common Mistakes

Best Practices

Hamesha Pipeline use karo — production ya real project mein manual approach avoid karo.
handle_unknown='ignore' OneHotEncoder mein lagao — naya category aane par error nahi ayega.
random_state=42 set karo — reproducible results ke liye (har baar same output).
Pipeline ko joblib se save karo: joblib.dump(pipeline, 'model.pkl') — ek file mein sab kuch.
R² score ke saath MAE bhi report karo — R² alone poori picture nahi deta.

Common Mistakes

Test data par fit_transform use karna — yeh data leakage hai! Test par sirf transform() use karo.
Scaler ka fit training se alag karna — naya scaler fit karoge toh mean/std alag honge.
Regression mein accuracy use karna — yeh bilkul wrong hai, MAE/MSE/R² use karo.
n_estimators bahut kam rakhna — 10-20 trees se underfitting ho sakti hai. 100-500 try karo.
Null values ignore karna — model fit hone ke baad bhi error aayega prediction mein.

Pipeline Save/Load — Production Tip

import joblib

# Save karo
joblib.dump(pipeline, 'housing_pipeline.pkl')

# Kisi bhi machine par load karo
loaded_pipeline = joblib.load('housing_pipeline.pkl')
result = loaded_pipeline.predict(new_data_df)

Results for "ml-beginners"

California Housing Price Prediction: Random Forest + Sklearn Pipeline