Cross-validation — k-fold: Foundations — Hinglish deep dive

5.6.5 · D1 · Coding › Machine Learning (Aerospace Applications) › Cross-validation — k-fold

Parent note Cross-validation — k-fold padhne se pehle, tumhe usmein istemaal hone wale har symbol ki understanding earn karni hogi. Yeh page unhe ek-ek karke introduce karta hai — har ek ke saath plain-words meaning, ek picture, aur reason ki yeh topic ko uski zaroorat kyun hai. Yahaan kuch bhi yeh nahin maanta ki tumne pehle yeh notation dekha hai.

1. Ek dataset $D$ , uska size $n$ , aur uske samples $(x_{i}, y_{i})$

Chaliye us line ke har piece ko left se right decode karte hain.

Curly braces ${\dots}$ ka matlab hai "cheezon ka ek collection" — yahaan, rows ka collection.
Har row ek pair hai jo round brackets mein likha hai $(x_{i}, y_{i})$ : ek input apne sahi answer se juda hua.
Neeche ka chhota number, subscript $i$ , bas ek row label hai — " $i$ -th row". Toh $x_{3}$ row 3 ka input hai.
Teen dots $\dots$ ka matlab hai "isi pattern mein aage badhte raho", last row $n$ tak count karte hue.

Parent ke aerospace example mein, $n = 1000$ hai aur ek row ek flight record hai: $x_{i}$ mein 50 sensor readings hain (temperature, vibration, pressure), aur $y_{i}$ ek single bit hai — kya engine 100 ghante ke andar fail hua ( $1$ ) ya nahin ( $0$ )?

2. Fold count $k$ aur fold sizes

Kyunki hum $n$ se mil chuke hain, dono ko alag rakho — yeh bahut alag roles play karte hain:

Kaun bada hai,

n

k

n

— tumhare paas bahut saari rows hain (

n = 1000

) jo kuch folds mein split hoti hain (

k = 5

Agar

n = 1000

aur

k = 5

ho, toh har fold kitni badi hogi?

\frac{n}{k} = \frac{1000}{5} = 200

rows.

Neeche ki picture woh clean case dikhati hai jahan chop evenly divide hota hai.

3. Folds $D_{i}$ aur "carve out" operations $\cup$ , $\cap$ , $∖$

Fold equation likhne se pehle, humein folds ko naam dena hoga.

Parent phir likhta hai $D = D_{1} \cup D_{2} \cup \dots \cup D_{k}$ . Teen naye symbols usme chhupe hain. Har ek rows ke groups ke baare mein baat karne ka ek tarika hai.

Ab parent ki line $D = D_{1} \cup D_{2} \cup \dots \cup D_{k}$ bas yeh kehti hai: "saare folds ko wapas jodne se poora dataset rebuild hota hai — kuch lose nahin, kuch add nahin."

Aur $D_{i} \cap D_{j} = \emptyset$ for $i \neq = j$ (padho " $i$ not equal to $j$ ") kehta hai: "do alag folds koi row share nahin karte." Yeh disjoint property hai, aur isliye ek row ek hi round mein kabhi trained-on aur tested-on dono nahin ho sakti.

Set-minus symbol parent ki training-set definition ko power deta hai: $D_{train}^{(i)} = D ∖ D_{i}$ Zor se padho: "round $i$ ka training set poora dataset hai jisme fold $i$ remove kar diya gaya hai." Upar superscript $(i)$ brackets mein bas ek round label hai — "round $i$ mein istemaal hua version", na ki koi power. (Upar number ke around brackets standard tarika hai yeh signal karne ka ki "yeh ek label hai, multiply mat karo.")

4. Learning algorithm $A$ aur trained model $M_{i}$

$A$ ko ek khaali brain samjho aur parentheses $A (\dots)$ ko "yeh data brain mein daalo" samjho. Bahar ek specific fitted model $M_{i}$ nikalta hai. Kyunki har round ek alag fold remove karta hai, har $M_{i}$ ek thoda alag model hai.

M_{i}

par subscript kyun?

Kyunki

k

alag trained models hain, ek per round —

M_{1}, M_{2}, \dots, M_{k}

5. Metric function aur score $Score_{i}$

Metric apne dono inputs ka kya use karta hai? Pehla, $M_{i}$ , guesses karta hai; doosra, $D_{test}^{(i)}$ , guess karne ke liye inputs aur compare karne ke liye hidden true answers dono supply karta hai. Parent do common metrics use karta hai:

Accuracy — guesses ka fraction jo correct hain. Simple, lekin jhutha jab ek class rare ho (ek 5%-failure dataset par hamesha "no failure" guess karna 95% deta hai).
F1 score — ek balanced number jo rare cases pakadne aur galat alarm na bajane dono ko reward karta hai. Prefer kiya jaata hai jab failures rare hoon.

Tumhe abhi F1 formula ki zaroorat nahin — bas jaano ki $Score_{i}$ ek honest grade per round hai, aur yahaan higher better hai.

6. Summation $\sum$ , mean, aur CV score

Yeh woh symbol hai jo sabse zyada beginners ko daraata hai. Ek baar unpack ho jaaye toh friendly hai.

Ab headline formula free mein nikal aata hai:

Average kyun, ek pick kyun nahin? Ek round ko ek easy test fold mil sakta hai aur luck se high score aa sakta hai. $k$ rounds average karne se good luck bad luck ko cancel karta hai, true performance ka ek steadier estimate deta hai.

7. Spread: standard deviation $σ_{CV}$

Average akela consistency chhupata hai. Parent ek spread bhi report karta hai.

Ise piece by piece decode karo, andar se bahar:

$Score_{i} - CV_{k}$ — ek grade average se kitna door hai (ek "miss").
$(\dots)^{2}$ — ise square karo, taaki negative misses aur positive misses dono positive distance count hon.
$\sum \dots$ — saare squared misses add karo.
$\frac{1}{k - 1}$ — divide karo (almost average miss; $k$ ki jagah $k - 1$ use karna ek standard small-sample correction hai).
— squaring undo karne ke liye square root lo, original units mein wapas aao.

Yeh pieces topic ko kaise feed karte hain

Is page par har cheez ek prerequisite arrow hai jo ek hi destination par point karta hai: ek single trustworthy number (plus ek wobble) ki model aisi data par kitna achha karega jo usne nahin dekha.

Aage kahan jaana hai

Parent itself: Cross-validation — k-fold
Folds se pehle, sabse simple idea: 5.6.01-Train-test-split-validation-set aur 5.6.02-Holdout-method
Rare classes ko folds mein balanced rakhna: 5.6.04-Stratified-sampling
$k$ ko $n$ tak push karna: 5.6.06-Leave-one-out-cross-validation-(LOOCV)
Kyun $k$ reliability affect karta hai: 5.6.10-Bias-variance-tradeoff-in-model-selection
Hinglish mein padhna prefer karo: 5.6.05 Cross-validation — k-fold (Hinglish)

Equipment checklist

Tum parent note ke liye ready ho jab tum bina dekhe in mein se har ek ka jawab de sako.

(x_{i}, y_{i})

kya stand karta hai?

Data ki row

i

— inputs

x_{i}

correct answer

y_{i}

ke saath paired.

n

aur

k

mein kya fark hai?

n

rows ki number hai;

k

folds ki number hai jinmein hum unhe kaatते hain.

Jab

n

k

se divisible nahin hota, har fold kitni badi hoti hai?

Ya toh

⌊ n / k ⌋

⌈ n / k ⌉

rows — zyada se zyada ek sample ka fark.

D_{i}

kya denote karta hai?

i

-th fold — piece number

i

mein rows ka bundle.

D_{i} \cap D_{j} = \emptyset

tumhe kya batata hai?

Alag folds koi row share nahin karte — woh disjoint hain, isliye koi row trained-on aur tested-on dono nahin hoti.

D ∖ D_{i}

ka matlab kya hai?

Poora dataset jisme fold

i

remove ho — yaani round

i

ka training set.

A

aur

M_{i}

mein kya fark hai?

A

khaali recipe hai;

M_{i}

woh model hai

A

ke baad jab woh round-

i

training data se seekh chuka ho.

Metric

kis tarah ki object hai?

Ek function jo ek model aur ek dataset leta hai aur ek number (ek scalar score) return karta hai.

\sum_{i = 1}^{k} Score_{i}

kya compute karta hai?

Saare

k

round grades ka sum:

Score_{1} + \dots + Score_{k}

Us sum se

CV_{k}

kaise nikaalte ho?

k

se divide karo — yeh

k

grades ka average hai.

Bada

σ_{CV}

tumhe kya warn karta hai?

Model ki performance bohot swing karti hai depending on which data woh dekhta hai — woh consistent nahin hai.

Recall Self-test: memory se poori pipeline rebuild karo

Paanch steps zor se bolo: (1) $D$ ko $k$ disjoint folds mein split karo; (2) har fold ke liye, $D ∖ D_{i}$ par train karo aur $D_{i}$ par test karo; (3) har round ko ek metric se score karo; (4) $CV_{k}$ paane ke liye scores average karo; (5) spread ke liye standard deviation $σ_{CV}$ lo. Agar tum yeh saare paanch kar sako, parent note kholo.