Lesson · Hinglish · ⏱ 8 min padhne mein
📏 Models ko measure karna: accuracy, precision aur recall
EnglishHinglish
🏠 Asli zindagi ki misaal
Airport security do tarah ki galtiyaan kar sakti hai: ek innocent passenger ko rokna (false alarm), ya asli khatre ko miss karna. Jo scanner kabhi beep hi na kare woh '99.99% accurate' hai, kyunki lagbhag koi khatra hota hi nahi, lekin bilkul bekaar hai! Precision poochta hai: jab beep hua, to kitni baar sahi tha? Recall poochta hai: saare asli khatron mein se kitne pakde?
💡 Aasaan shabdon mein
Accuracy, yaani sahi predictions ka hissa, tab misleading hoti hai jab ek class rare ho, jaise fraud ya bimaari. Hamesha ek baseline se compare karo, jaise 'hamesha sabse common class predict karo'. Precision = sahi positives ÷ jitne aapne flag kiye. High precision matlab kam false alarms. Recall = sahi positives ÷ saare asli positives. High recall matlab kam miss. F1 dono ko combine karta hai. Kaunsa zyaada zaroori hai, yeh har galti ki keemat pe depend karta hai: cancer screening mein recall critical hai. Spam filter mein precision zaroori hai taaki asli emails chhup na jaayein. Regression ke liye MAE (average absolute error) jaise errors use karo.
🎬 Dekho kaise kaam karta hai
1,000 payments mein se 10 fraud hain. Model 20 flag karta hai, jinmein 8 asli hain. Precision 'flagged' row mein dekhta hai, recall 'actually fraud' column mein.
🪜 Step by step
- Confusion matrix banao: true positives, false positives, false negatives, true negatives.
- Precision = TP ÷ (TP + FP) aur recall = TP ÷ (TP + FN) nikalo.
- Business se poochho kaunsi galti mehengi hai: miss hua fraud, ya block hua asli customer?
- Threshold khiskao (0.5 pathar ki lakeer nahi) taaki zaroori precision aur recall ka balance mile.
- Chuna hua metric test set pe baseline ke saath compare karke report karo.
🔑 Zaroori shabd
- Precision
- Jitna flag kiya, usmein se kitna sach mein positive tha.
- Recall
- Saare asli positives mein se kitne pakde gaye.
- Confusion matrix
- Sahi aur galat yes/no predictions ki 2×2 table.
👀 Example dekho
1,000 transactions, 10 are fraud.
Model flags 20; 8 of them are real fraud.
Precision = 8 / 20 = 40% (many false alarms)
Recall = 8 / 10 = 80% (caught most fraud)
Accuracy = (8 + 978) / 1000 = 98.6%
Baseline 'never fraud' accuracy = 99% ← beats the model on accuracy!Accuracy kahani chhupa deti hai. Precision aur recall use dikhaate hain.
⚠️ Aam galtiyaan jinse bachna hai
- Imbalanced data pe high accuracy ka jashn manana. Dekho 'hamesha nahi' baseline kitna score karta hai.
- Aisa metric chunna jiski business ko parwah nahi. Translate karo: 'hum 10 mein se 8 fraud pakadte hain aur 500 mein se 1 customer galti se block karte hain'.
✅ Jaldi se check karo
Cancer screening model ke liye aam taur pe kaunsi galti zyaada buri hai?
- Dono barabar theek hain
- Low recall: asli cases miss karna
- Low precision: kuch false alarms
Jawaab dekho
B. Low recall: asli cases miss karna: Miss hua cancer case jaan le sakta hai, jabki false alarm ka matlab bas ek aur test. Isliye recall sabse zaroori hai.
🙋 Seekhne wale yeh poochhte hain
ROC AUC aur PR AUC kya hain?
Dono saare thresholds pe performance summarise karte hain. Rare positives ke liye PR AUC zyaada kaam ka hai kyunki ROC AUC tab bhi accha dikh sakta hai jab precision kharab ho.
Regression ke liye kaunse metrics?
MAE (error ka average size), RMSE (bade errors ko zyaada saza) aur MAPE (percent mein error). Woh chuno jo galtiyon ke asar se match kare.
📌 Yaad rakho
Baseline ko beat karo, aur har galti ki keemat dekhkar precision vs recall chuno.
📖 Is lesson ke shabd (AI dictionary)
Poora course, aapke roadmap ke hisaab se
Progress tracking, quick checks aur step-by-step order ke saath. Yeh lesson in courses ka hissa hai:
AI Foundations Path · ML Engineer · AI Research Engineer · AI Product Manager
Course mein seekho