Regression Algorithms: Linear se Ridge/Lasso tak

Jab humein kisi continuous number ko predict karna ho (jaise ghar ki price ya stock market), toh hum Regression use karte hain. Lekin sirf ek "Straight Line" khinchna kafi nahi hota. Real-world data aksar "Overfit" ho jata hai ya phir bahut "Tedha" (Non-linear) hota hai.

1. Linear Regression: The Baseline

Linear Regression ek seedhi line dhoondhta hai ($y = mx + c$).

• Cost Function: Model ye dhoondhta hai ki "Galti" (Error) sabse kam kahan hai.
• Gradient Descent: Ye wo "Dheere-dheere utarne" ka rasta hai jisse model apne weights ko optimize karta hai taaki error minimum ho.

2. Polynomial Regression: Curved Trends

Har cheez line mein nahi chalti.

• Agar aapka data "S" shape ya curve mein hai, toh hum features ki power ($x^2, x^3$) badha dete hain.
• Ise Polynomial Regression kehte hain. Par dhyan rahe, zaroorat se zyada power model ko "Ziddi" (Overfit) bana sakti hai.

3. Regularization: Ridge aur Lasso

Jab model training data ko "Rat" (Memorize) leta hai, toh wo naye data par fail ho jata hai. Isse bachne ke liye hum Penalty lagate hain.

• Ridge (L2): Ye weights ko chhota karta hai par zero nahi karta.
• Lasso (L1): Ye faltu features ke weights ko Zero kar deta hai (Feature Selection).

4. Evaluation: R-squared vs Adjusted R-squared

• R-squared ($R^2$): Batata hai ki model kitna accha hai (0 se 1 tak).
• Adjusted $R^2$: Ye tab use hota hai jab hamare paas bahut saare features hon. Ye faltu features add karne par score ko "Punish" karta hai.

5. Summary Table: Regression Types

Regularization Deep-Dive: Ridge (L2) vs Lasso (L1) vs ElasticNet

Standard Linear Regression (Ordinary Least Squares) multi-collinear data aur outliers par severely overfit karta hai:

1. Ridge Regression (L2 Penalty): Cost function mein lambda * sum(weights^2) add hota hai. Ye weights ko zero ke paas shrink karta hai lekin exact zero kabhi nahi karta. Jab saare features important hon aur unme collinearity ho, tab Ridge best hai.

2. Lasso Regression (L1 Penalty): Cost function mein lambda * sum(|weights|) add hota hai. Ye less important features ke weights ko exact 0 bana deta hai — effectively performing automatic Feature Selection.

3. ElasticNet: L1 aur L2 ka combination hai. Jab features correlated hote hain, toh Lasso randomly kisi ek feature ko chunta hai aur baaki ko zero kar deta hai; ElasticNet dono ko group mein retain karta hai.

from sklearn.linear_model import Ridge, Lasso, ElasticNet

ridge = Ridge(alpha=1.0)
lasso = Lasso(alpha=0.1)
elastic = ElasticNet(alpha=0.1, l1_ratio=0.5)

📚 Sambandhit Tutorials & Agle Kadam

Agar aap is topic par apni pakad mazboot karna chahte hain, toh in dedicated guides ko zaroor padhein:

• 🔗 Classification Algorithms: Decisions lene ka dimaag — Machine Learning practical guide.
• 🔗 Loss Functions: AI ki galti naapne ka scale — Deep Learning practical guide.
• 🔗 Bias vs Variance Tradeoff: ML Models ka Core Balance (Python Code) — Machine Learning practical guide.
• 🔗 Machine Learning: Zero se Hero tak (2026 Edition) — Machine Learning practical guide.

FAQs

1. Gradient Descent mein "Learning Rate" kya hai? Ye wo "Step size" hai jisse model seekhta hai. Agar ye bahut bada ho toh model rasta bhatak jayega, agar bahut chhota ho toh training mein sadiyan lag jayengi.

2. Ridge aur Lasso mein se kaunsa chunu? Agar aapko lagta hai ki aapke saare features important hain, toh Ridge use karein. Agar aapko lagta hai ki kuch hi features kaam ke hain, toh Lasso use karein.

3. "Multicollinearity" kya hoti hai? Jab aapke do inputs aapas mein jude hon (e.g., Temperature in Celsius and Fahrenheit). Ye regression model ko confuse kar dete hain.

4. RMSE aur MAE mein kya fark hai? RMSE badi galtiyon ko zyada "Saza" (Square) deta hai. MAE sirf average error batata hai. AI mein RMSE zyada popular hai.

Regression Machine Learning ka "Pehla Kadam" hai. Ise samajh kar aap kisi bhi numerical trend ko predict kar sakte hain! 📈