Machine Learning Comprehensive Notes

Unit I: Introduction to Machine Learning and Foundations

Machine Learning (ML) enables computers to learn patterns from data and make predictions without
being explicitly programmed. It is a subfield of Artificial Intelligence (AI) where algorithms improve their
performance with experience 1 2 . In other words, ML algorithms build a mathematical model
(hypothesis function) that maps inputs to outputs based on example data, then generalize to predict
outcomes on new data 1 2 . For instance, an email spam filter learns from past labeled emails to predict
whether new emails are spam. In ML we train a model on historical data so it can make accurate decisions
or classifications on unseen data.

Hinglish: Machine Learning ka matlab hai computer ko data se seekhna aur bina explicitly code kare
decisions ya predictions karna 2 . ML algorithms ek hypothesis function (jaise regression line ya neural
network) seekhti hain jo input features se output predict karti hai. Jaise agar humne email ke spam/ham
labels diye, ML algorithm ko past emails se training de ke, woh naye emails ko spam ya non-spam predict
kar sakta hai 2 .

Udāharan (Hinglish):
- Email Spam Detection: Kuch logon ko har roz hazaaro emails milte hain; ML model train karne se 85% se
upar accuracy ke saath spam emails block kiye ja sakte hain.
- Credit Approval: Bank ka ML model customer ke credit score aur income dekh ke “approve” ya “reject”
decide kar sakta hai.

Figure: Machine Learning Life Cycle phases. Machine Learning ka life cycle ek iterative process hai jisme pehle
problem define kiya jata hai, phir data preparation (cleaning, exploration, feature engineering) ki jati hai,
model development me algorithm choose karke training ki jati hai, uske baad model deployment me real
system me daal ke use kiya jata hai, aur monitoring & maintenance phase me model ki performance
dikhte data ke saath check ki jati hai. Misal ke taur par, ek ML project me data collect karna, missing values
fill karna, features select karna, fir train-test split karke model banana, aur system me deploy kar ke results
monitor karne tak sab steps hote hain. Ye sab steps complete hone ke baad bhi process repeat ho sakta hai
jaha naya data aane par model update karte rahte hain.

Hinglish: ML life cycle me pehle business problem samjha jaata hai, phir data ki quality check aur cleaning
hoti hai. Phir algorithms chuni jaati hain, model train aur validate hota hai, aur jab kaam theek lage to
production me deploy kiya jata hai. Model ki accuracy time ke saath gir sakti hai to monitoring aur
retraining jaruri hai.

Udāharan (Hinglish):
- ML Pipeline Example: Sochiye ek problem hai “dog photo ya cat photo”. Life cycle ke steps: (1) Dog-cat
photos collect karo, (2) Unko resize/normalize karo (3) CNN model train karo (4) Accuracy check karo (5)
Agar okay ho to app me dal ke har nayi photo test karo (6) Time-time pe model ko retrain karte raho.

1
Types of Learning Systems

• Supervised Learning: Training data me features ke saath labels (target values) bhi hote hain. Model
un labels ke hisaab se predict karna seekhta hai (jaise classification/regression) 3 . Udaharan: Agar
har fruit ke saath “apple” ya “banana” label diya ho, supervised model naye fruits ko sahi label
karega.

• Unsupervised Learning: Training data me labels nahi hote; model aise tasks karta hai jahan data ke
patterns ya structure khud discover karna hota hai, jaise clustering ya dimensionality reduction
3 . Udāharan: Agar sirf fruit images hain bina labels ke, unsupervised algorithm similar fruits ko ek
saath group karega.

• Semi-supervised Learning: Thode labeled aur bahut saare unlabeled examples ke saath learning
hoti hai. Yeh beech ka mixture hai jahan model kuch labels se guidance paata hai par adhikansh data
unlabeled rehta hai.

• Reinforcement Learning: Yahan model koi actions leta hai aur environment se reward/penalty milta
hai, jisse wo apni policy improve karta hai. (Not covered in detail here.)

• Batch vs Online Learning: In Batch Learning, model ko pehle se puri training dataset di jati hai aur
ek saath train kiya jata hai; naye data ke liye model ko phir se retrain karna padta hai. Online
Learning me model initial training ke baad continuously updates ke saath naye data ke saath sikhta
rahta hai. Udāharan: Spam filter ko har roz naye emails se seekhna ho, to online learning useful hai.

• Instance-based vs Model-based Learning: Instance-based (memory-based) learning me algorithm

training examples ko memory me store karke, naye examples ke liye in stored instances se directly
similar ones dhundkar output deta hai (jaise k-NN). Model-based learning me algorithm ek general
model (jaise linear regression ya neural network) banata hai, uske parameters (weights) optimize
karke future data predict karta hai. Instance-based simple aur flexible hai, par memory-intensive;
model-based compress training info into fixed parameters.

Hinglish: ML ke mukhya types hai: Supervised jaha labeled data se train hota hai 3 (jaise image ke saath
uska naam), Unsupervised jaha labels nahi hote aur algorithm khud clusters ya patterns dhundhta hai 3 .
Batch learning me model ek baar me poora data use karke train hota hai, Online learning me naya data aane
par model update hota rehta hai. Instance-based (jaise k-NN) me examples store karke unke basis pe naye
cases solve karta hai; Model-based me ek formula/parameters banta hai jise optimize karke prediction hoti
hai.

Udāharan (Hinglish):
- Supervised Example: Aap ke paas fruits ke photos aur unke names (apple, orange) hain; model train
karke naye fruits ko pehchaan lo.
- Unsupervised Example: Bank transactions ko labels nahi diye; clustering se similar kharche waale
customers group karo.
- Batch vs Online: Agar medical sensor ka data baar-baar aa raha hai, to online learning model har naya
reading pe update karega; batch me sab readings ikattha kar train karenge.
- Instance vs Model: k-NN (instance-based) ke liye poore database jaruri hai; jabki linear regression (model-
based) me kuch weights find hote hain.

2
Scope, Limitations, and Challenges of Machine Learning

Machine Learning aaj har industry me bahut important hai. Scope: ML models health care (disease
prediction), finance (fraud detection), e-commerce (recommendations), image processing, speech
recognition, aur climate science jaisi fields me use hote hain 4 . Jaise 2024 me ML healthcare me
diagnoses aur treatments me, finance me algo-trading me, marketing me customer segmentation me use
ho raha hai 4 .

Limitations: ML powerful hai lekin kuch limitations bhi hain. ML ko bahut saara high-quality data
chahiye; agar data kam ho ya biased ho to model achha train nahi karega. Agar training data representative
na ho (sampling bias), model real-world me galat predictions dega. Kuch features irrelevant hain to model
confuse ho sakta hai. Overfitting/underfitting bhi challenge hai: overfitting me model training data pe bhi
naach raha hota hai par naye data pe fail, underfitting me model simple ho ke pattern nahi pakad pata 5
6 . ML ke kuch tasks (jaise commonsense reasoning) abhi tak humans se achha nahi kiya ja sakta.

Hinglish: ML bahut scope wala hai — healthcare me bimari predict karna, mobile apps me face recognition,
India me bhi ML startups bohot badh rahe hain 4 . Par ML ki limitations bhi hain: agar training data quality
kharaab ho ya bahut kam ho to model sahi se nahi seekh paata. Bhari models ko samjhana mushkil hai
(interpretability), aur compute power ki bhi zaroorat padti hai. Overfitting (model sirf training ache se
memorize karke naye pe fail) aur underfitting (model bahut simple hona) common problems hain 5 6 .

Udāharan (Hinglish):
- Data Quality: Agar kisi zilla ke ganna anaj ka data adhoora hai, to ML model poore Bharat ke anaz trend
predict nahi kar payega.
- Representativeness: Agar sirf urban users ke data pe recommenders train huye, rural users ke liye kaam
nahi karenge.
- Overfitting Example: Agar ek decision tree ne training set ke har record me perfect fit kiya (har branch
me ek hi example), to naye data pe galat classification ho sakta hai 6 .
- Scope Example: Gaon me laddu recipe ke pictures pe ML train karke Laddu pehchanne wala app bana
sakte hain (digital India initiative).

Data Visualization and Hypothesis Concepts

Data Visualization: ML me data ko visualize karna (charts, graphs, plots) bahut zaroori hai. Visualization se
data ke trends, patterns, outliers jaldi pakde ja sakte hain 7 . Achhi exploratory data analysis aur charts se
ML process fast aur accurate banta hai 7 . Udāharan: Agar aap student ke marks vs study hours ko scatter
plot me dekhen, to correlation dikhegi jo model building me help karegi 7 .

Hinglish: Data ko line charts, bar graphs ya scatter plots me plot karke hum clearly dekh sakte hain ki
features ka distribution kaisa hai. Ye ML process ko aasan banata hai kyunki important features ya patterns
visually samajh aa jaate hain 7 .

Udāharan (Hinglish):
- Example: Ek bank apne loan applicants ke age aur annual income ka scatter plot banaye; agar data me
cluster dekhe to pata chal sakta hai kaunse age-income group me loan default zyada ho raha hai.

3
Hypothesis Function and Testing: ML me hypothesis function (h_θ(x)) ka matlab hota hai
model ka formula jo input x se output predict karta hai 8 . Ye aapke model ka roop hai (jaise linear
regression h(x)=w₀+w₁x). Hypothesis testing ka matlab hai hum apne model ke
assumptions ko data se check karein — jaise statistical tests me hum null hypothesis test karte hain ki
model ka prediction chance se alag hai ya nahi. Supervised learning me hum training ke baad apne model
ki prediction accuracy test data pe measure karte hain, jise kabhi hypothesis testing bhi bola jaata hai.

Hinglish: Hypothesis function matlab wo mathematical function (jaise linear ya sigmoid) jise model ne
seekha hai inputs se output nikalne ke liye 8 . Model testing me hum check karte hain ki ye function naye
data pe sahi output de raha hai ya nahi. Aksar ham log model ki accuracy, error metrics nikal ke dekhte hain
ki assumption (hypothesis) data ko theek fit karta hai ki nahi.

Udāharan (Hinglish):
- Example: Linear regression me hypothesis function h(x)=a+bx; testing me dekhenge ki actual y values se
in predictions ka difference (error) kitna hai.

Data Pre-processing, Augmentation, and Normalization

Data Pre-processing: Raw data me aksar missing values, duplicates, ya irrelevant entries hoti hain. Data
pre-processing me hum missing values ko handle, outliers ko detect, irrelevant features hataye, aur categorical
data ko numeric me convert karte hain. Tutorialspoint ke mutabik, data preparation ke phases me data
exploration (visualization, summary), feature engineering (naye features banana) aur feature selection
(relevant features chunna) shamil hai. Example: Kaggle dataset me missing entries ko mean/median se fill
karna, ya text ko one-hot encoding me badalna.

Hinglish: Data pre-processing me hum data ko saaf-suthra banate hain. Matlab missing values fill karte,
duplicate records nikalte, galat entries sudhaarte hain, aur jo features model ke kaam ke nahi unko hata
dete hain. Ye kadam model accuracy ke liye bahut important hai.

Udāharan (Hinglish):
- Agar temperature sensor me kuch readings missing hain, to unhe average se fill karo. Agar address field
bahut noisy hai to use hata do. Ye sab preprocessing steps hain.

Data Augmentation: Jab dataset chhota ya imbalance ho, tab data augmentation techniques use hoti hain.
Ye pre-existing data se naye synthetic examples banati hai taki model zyada robust ho. Jaise images ke liye
rotation, flip, zoom, noise add karna 9 . IBM kehta hai ki data augmentation model ko better generalize
karne me madad karta hai, aur overfitting reduce karta hai 9 10 . Example: Agar humare paas 100 dog
images hain, to unka mirror image bana ke 100 aur bana sakte hain.

Hinglish: Data augmentation ka matlab hai existing data pe chhote transformations (jaise rotate, flip ya add
noise) karke naye training examples create karna 9 . Isse humara model har tarah ke data pe train ho jata
hai aur zyada acche se generalize karta hai. Jaise image classifier me ek dog ki photo rotate karke alag
angle se dog dikhaya jaata hai.

Udāharan (Hinglish):
- Image Augmentation Example: Agar sirf 10 handwritten digits hain, to unko thoda jhukake, zoom karke,
ya brightness badha ke naye samples banayo taaki neural network aur acche se train ho jaye.

4
Normalization and Scaling: Machine Learning me normalization ya feature scaling bohot zaroori hai jab
features ki values alag-alag range me ho. Normalization me numeric features ko ek common range (jaise
[0,1]) me map kiya jata hai 11 . GfG ke hisaab se, normalization se ensure hota hai ki har feature ka same
scale rahe, taki high-magnitude feature model ko dominate na kare 11 . Isse distance-based algorithms
(jaise k-NN, SVM) me accuracy aur convergence improve hoti hai 11 .

Hinglish: Normalization me hum features ke values ko chhota scale pe le aate hain (jaise 0 se 1), jisse ek
feature dusre pe bhaari na padhe 11 . Isse model training fast hoti hai aur sab features barabar contribute
karte hain.

Udāharan (Hinglish):
- Agar ek feature 0–1 range me hai aur dusra 0–1000 range me hai, to pehle ko 0–1 me aur dusre ko 0–1 me
scale karo (min-max scaling) taki model training me balance ho.

Bias–Variance Tradeoff and AI–ML–DL–DS Relationship

Bias–Variance Tradeoff: ML me error do parts ka hota hai: Bias aur Variance. High bias matlab model bohot
simple hai (underfitting), jisse training aur test error dono badh jate hain 5 . High variance matlab model
bohot complex hai (overfitting), jisse training error to kam hota hai par test error jyada 6 . In dono me
balance banana hi bias–variance tradeoff hai 12 . Neeche wale figure me different models ka bias-variance
illustrate hua hai:

Figure: Bias–variance decomposition – simple model (high bias, low variance) se lekar complex model (low bias,
high variance) tak tradeoff. 5 6 .

In tradeoff ke mutabiq, hum model complexity aisi choose karte hain ki total error (Bias² + Variance +
irreducible error) kam ho 12 . Regularization techniques is tradeoff ko manage karne me madad karti hain.

Hinglish: Bias matlab model ka training-data se bhi mismatch hona (underfitting), variance matlab model
training set ke chote changes se bahut farak dikhana (overfitting) 5 6 . Agar model bahut simple rakha
to high bias hoga, agar bahut complex to high variance hoga. Hum dono ke beech ka sahi santulan (trade-
off) chunke error minimize karte hain 12 .

Udāharan (Hinglish):
- High Bias Example: Ek linear model lena jab data non-linear ho – training aur testing dono me high error
(underfitting) 5 .
- High Variance Example: Ek bohot gehra decision tree banana jisne training ke sab points perfectly fit kiye
– test par galat answer dera hai (overfitting) 6 .

Figure: Relationship between AI, ML, DL, and Data Science 13 . Ye diagram dikhata hai ki Artificial Intelligence
sabse bada term hai; uske andar Machine Learning (ML) aata hai, jisme se Deep Learning (DL) ek subfield
hai (neural networks ka). Data Science ek alag field hai jo data analysis aur ML/DL tools ka use karke
insights nikalta hai 13 . Deep Learning (jaise CNN, RNN) ML ka hissa hai, aur ML khud AI ka part hai. Data
Science me ML/DL ke saath data engineering, visualization aur statistics shamil hote hain.

5
Hinglish: AI me ML aur ML me DL included hai 13 . Matlab DL (deep neural networks) ML ka subset hai, aur
ML AI ka subset. Data Science ek broader kshetra hai jisme ML/DL ko data analysis ke saath milake use kiya
jata hai.

Udāharan (Hinglish):
- Example: AI se hum smart machines banate hain; uske liye ML algorithm (like spam detection, speech
recognition) use kar sakte hain; Deep Learning me hum complex neural networks (jaise image recognition)
apply karte hain. Data Science me companies ye sab data ke analysis ke liye istemal karte hain (marketing
analysis, customer segmentation, etc.).

Unit II: Clustering in Machine Learning

Clustering is unsupervised learning ka ek method jisme similar data points ko ek saath group kiya jata hai
(clusters) bina kisi label ke 14 . Maksad hai ki ek cluster ke andar wale points ek dusre se zyada similar hon
aur dusre clusters se alag hon. Clustering se hum data ke hidden structure ko samajh sakte hain, jaise
customer segmentation ya image segmentation.

Hinglish: Clustering me hum data points ko group banate hain jise cluster kehte hain. Har cluster me ek jaisi
cheezein hoti hain. Yeh bilkul waise hi hai jaise same gharon wale areas ko ek cluster me rakhna. Data ke
clusters batate hain ki kaunse data points ek jaise hain.

Udāharan (Hinglish):
- Customer Segmentation: Company apne customers ke shopping habits ke aadhar par clusters banati
hai, jaise luxury shoppers, budget shoppers, etc.
- Document Clustering: News articles ko topics ke hisaab se bina labels ke cluster kiya ja sakta hai (sports
vs politics).

Types of Clustering Methods

• Partitioning Clustering (Centroid-based): Data ko k clusters me divide karta hai jahan har cluster
ek center (mean/centroid) se represent hota hai. K-means aur K-medoids is ke examples hain 15 . In
methods me aapko pehle se clusters ki sankhya (k) bataani padti hai. K-means jaisa algorithm
randomly k initial centers leta hai aur phir data points ko unke nikat centroid ke hisaab se assign
karke iteratively centers adjust karta hai. Table 1 me partitioning vs hierarchical vs fuzzy clustering
compare kiya gaya hai.

Partitioning Clustering Description

Partitioning method jaise k-means, data ko pre-specified number

Uses k centroids to form k
(k) of clusters me divide karta hai 15 . Har cluster ko ek centroid
clusters 15
represent karta hai.

6
 Partitioning Clustering Description

Assumes spherical (simple) Doosre algorithms ki tarah outliers ko handle karne me mehsoos
cluster shapes; efficient on large nahi, ek-dusre cluster ke centers se door points iss tarah ki cluster
datasets me add ho sakte hain.

• Distribution Model-Based Clustering: Data ko probability distributions (jaise Gaussians) ke mixture

se assume karta hai 16 . Udāharan: Gaussian Mixture Models (GMM) har cluster ko ek Gaussian
distribution ke roop me model karta hai. EM (Expectation-Maximization) algorithm se in Gaussian
parameters (mean, covariance, weights) estimate kiye jate hain 16 17 . Essentially, har point ke
cluster membership probabilities calculate kiye jate hain (soft assignments).

Hinglish: Distribution-based clustering me data ko kuch probability distributions se bana hua assume karte
hain. Example: GMM me data ko kuch Gaussians ka mixture maan ke clusters bante hain, EM algorithm se
parameters (mean, variance) find hota hai.

• Hierarchical Clustering: Data ke clusters ko tree-structure (dendrogram) me banata hai. Isme

agglomerative (bottom-up) approach me har data point apna cluster shuru me hota hai, phir step-
by-step nearest clusters merge hote hain; divisive (top-down) me saare points ek cluster se start
hote hain aur split hote hain. Koi pre-set k number nahi batana padta; dendrogram decider hota hai
kitne clusters final honge 18 . Neeche Figure me hierarchical clustering ka ek dendrogram diya hai.

Figure: Example hierarchical clustering dendrogram (complete-linkage). Har merge ek new cluster darshata hai
18 .

Hinglish: Hierarchical clustering me hum points ko tree jaisa jodte hain. Upar wale figure me sabse niche
single data se start karke groups bante gaye. Aap dendrogram cut karke desired cluster count decide kar
sakte hain.

• Fuzzy (Soft) Clustering: Har data point sirf ek cluster ka nahi, balki multiple clusters me
membership levels de sakta hai (0–1 tak weight) 19 . Common algorithm Fuzzy C-Means hai, jahan
har point ke liye probabilities nikalte hue centroid update karte hain. Fuzzy clustering me
overlapping clusters allowed hote hain 19 .

Hinglish: Fuzzy clustering me ek point ek se zyada clusters me partially belong kar sakta hai. Jaise market
segmentation me ek customer ko multiple segments me weight assign kar sakte hain.

• BIRCH (Balanced Iterative Reducing and Clustering using Hierarchies): BIRCH large dataset ke
liye incremental clustering technique hai 20 . Ye data ke liye ek compact summary (Clustering
Feature Tree) banata hai jo large dataset ko bahut chhote summaries me convert karta hai, phir un
summaries par clustering apply hoti hai 20 . Isse large datasets efficient tarah se cluster kiye ja sakte
hain.

• CURE (Clustering Using REpresentatives): CURE hierarchical algorithm hai jo her cluster ke liye
multiple representative points use karta hai, na ki sirf ek centroid 21 . Ye algorithm clusters ki shapes
ko acche se capture karta hai aur outliers ko eliminate karne me madad karta hai 21 .
Representative points ko pick karke cluster boundaries more robust bante hain.

7
Hinglish: BIRCH ek tarah se large data ke summary se clusters banata hai 20 , CURE me har cluster ke liye
kuch representative points liye jate hain jisse complex shapes aur outliers handle hote hain 21 .

Gaussian Mixture Models and Expectation-Maximization

Gaussian Mixture Model (GMM): Clustering me GMM ek probabilistic approach hai jahan data ko multiple
Gaussian distributions (normal curves) ka mixture maan ke model kiya jata hai 22 . Har Gaussian
component ek cluster ko represent karta hai. GMM me har cluster ka apna mean (centroid) aur covariance
hota hai, aur points ko soft probabilities se assign kiya jata hai (each point ka probability nikalte hain ki wo
har Gaussian se belong karta hai).

Expectation-Maximization (EM): GMM parameters (means, covariances, weights) ko find karne ke liye EM
algorithm use hota hai. EM iterative hai:
1. E-step: Current model parameters ke base par, har data point ke liye har Gaussian ke under likelihood/
probability calculate karo (soft assignments).
2. M-step: In probabilistic assignments ko use karke Gaussian components ke parameters (mean,
covariance, mixing weights) update karo.
Ye steps alternate karke converge hone tak repeat karte hain 17 . Isse final GMM model milta hai jo data ko
best explain karta hai.

Hinglish: GMM me hum sochte hain ki data kuch number ke Gaussian groups me banta hai 22 . EM se hum
un group ke parameters nikalte hain: pehle guess karke probability banate hain, phir update karke betar
estimates nikalte hain 17 .

Udāharan (Hinglish):
- GMM Example: Agar aap city ke temperature data ko cluster kar rahe hain, to shayad subah-dopehar-
sham ke hisaab se 3 Gaussian curves fit kar sakte hain.
- EM Example: Kisi chemical particle mix me, agar teen types ka signal hai, to EM se un teen Gaussian
signals ke peaks (mean) aur width (variance) find kar sakte hain.

Parameter Estimation – MLE and MAP

Maximum Likelihood Estimation (MLE): MLE model ke parameters ko is tarah choose karta hai ki
observed training data ka probability (likelihood) maximum ho jaye. Asan shabdon me, “jis parameters ke
saath data sabse zyada plausible lagta hai, wahi MLE estimate hota hai.”

Maximum A Posteriori (MAP) Estimation: MAP bhi data ko maximize karta hai, lekin ek prior (a priori
belief) ko include karke. MAP me hum parameters ke liye ek prior probability assume karte hain, phir
posterior (likelihood × prior) maximize karte hain. Isse overfitting kam hoti hai aur parameters par control
rehta hai. MAP MLE ka regularized form hai 23 :

Posterior ∝ Likelihood × Prior, so MAP maximizes P(parameters|data), jisme prior term

regularization jaisa kaam karta hai 23 .

Hinglish: MLE me hum bas dekhte hain ki training data given parameters ke saath sabse zyada aisa lagta
hai. MAP me hum parameters ke baare me pehle se thoda knowledge (prior) maan ke, parameters nikalte
hain. MAP ek tarah se priors lagakar training ko thoda rokta hai, taaki model bahut complex na ho 23 .

8
Udāharan (Hinglish):
- MLE Example: Normally assume karte hain ki errors Gaussian hain aur phir best-fitting line aisi nikalo ki
squared errors minimized ho (ye linear regression me equivalent hai).
- MAP Example: Agar hamne pehle se jaana hai ki weight decay (small weights) chaiye to MAP me “smaller
weights” ko thoda preference denge priors se.

Unit III: Classification Algorithms and Evaluation

Classification algorithms me hum labeled data se seekhkar categories predict karte hain. Neeche major
algorithms aur unke features diye gaye hain:

Logistic Regression

English: Logistic Regression supervised classification algorithm hai jo continuous input features ke basis pe
probability predict karta hai ki ek example positive class me aaega ya nahi 24 . Ye model ek logistic
(sigmoid) function use karke linear combination of inputs ko [0,1] range me map karta hai. Typically
binary classification ke liye use hota hai (jaise yes/no), magar multiclass ke liye bhi extend kiya ja sakta hai.
Logistic regression me hum parameters (weights) ko train set pe optimize karte hain taaki predicted
probabilities actual labels se match kare 24 .

Hinglish: Logistic Regression me input features ko ek weighted sum (like w0 + w1 x1 + …) ke through ek

sigmoid function (1/(1+e^(-z))) me daalte hain, jisse output probability milti hai. Ye probability bataati hai ki
example kis class ka ho sakta hai. Example: Agar output 0.8 aaya to 80% chance usko “spam” label milega.
Logistic regression mostly binary classification me istemal hota hai 24 .

Udāharan (Hinglish):
- Example: Email spam classifier: input features (jaise subject me “offer” word count, sender address
domain, etc.), logistic model train karke probability compute karta hai; fir decide karta hai spam/non-spam.

Decision Tree Classification

English: Decision Tree algorithm supervised learning me tree-like model banata hai jisme internal nodes
features par decisions (questions), branches outcomes, aur leaves final class labels hoti hain 25 . Ye
recursive partitioning karta hai: har node par sabse informative feature choose karke data split kar deta hai,
taki child nodes me data zyada homogeneous (same class) ho. For example, agar root node par “Income >
50k?” question hota hai, left branch high-income logs aur right branch low-income. Har leaf ek class (jaise
Yes/No) assign karta hai 25 . Decision trees interpret karne me aasaan hain (flow-chart jaise) aur
categorical/continuous features dono handle karte hain.

Figure: Example Decision Tree for predicting purchase (Yes/No) based on customer attributes 25 . In figure, root
node me “Income > 50k?” test hai, uske aage age threshold test hai, aur leaves me final prediction (Yes/No)
likha hai. Har path ek rule represent karta hai (jaise Income>50k and Age<30 → Likely Yes).

Hinglish: Decision Tree bilkul ek flowchart hota hai. Upar se root node start hota hai jahan feature check
hota hai, phir uske result se branches me data divide hota hai, aur ant me leaves me class label milta hai. Ye

9
algorithm easily samajh aa jata hai kyunki decisions “agar income zyada toh… warna…” jaise likha hua milta
hai.

Udāharan (Hinglish):
- Example: Loan approval: Tree me pehla question ho sakta hai “Credit Score > 700?”, agar haan to agla
“Existing debt < threshold?”, and so on. Leaves me “Approve” ya “Reject” diya hoga.

Artificial Neural Network (ANN)

English: Neural Network supervised model hota hai jo human brain ke tarah layers of connected neurons
(nodes) se bana hota hai. Yeh input layer, ek ya adhik hidden layers, aur output layer me divide hota hai. Har
connection ek weight rakhta hai. During training, network inputs ko hidden layers se forward propagate
karke output generate karta hai, phir backpropagation se errors kam karte hue weights update karta hai.
Deep Learning networks (jaise CNN, RNN) many hidden layers ke saath complex patterns (images, speech,
language) ko learn karte hain. Neural networks non-linear problems me bahut powerful hain.

Figure: Example Neural Network architecture with input, hidden, and output layers. Jaise figure me dikhaya hai,
input layer me features (x1, x2,…), hidden layer me neurons (y1, y2, …), aur output layer me predictions
(e.g., class scores) hain.

Hinglish: Neural Network ek multi-layer model hai. Input se start karke data hidden layers se jaata hai jaha
non-linear transformations hote hain, aur phir output layer class probabilities ya values deti hai. Ye weights
ke network hai jo training me adjust hote hain. Bahut complex patterns ko seekhne ke liye (jaise image
recognition) use hota hai.

Udāharan (Hinglish):
- Example: Image classifier: Input layer me image ke pixel values, hidden layers convolutional filters aur
activation functions se features nikalte, output me “cat vs dog vs horse” probabilities nikalte.

K-Nearest Neighbors (K-NN)

English: K-NN ek simple instance-based (non-parametric) algorithm hai. Supervised classification/

regression ke liye use hota hai. Naye point ki class predict karne ke liye K-NN training data me uske k sabse
nazdeek (usually Euclidean distance se) neighbors dekhta hai, aur un neighbors ka majority vote
(classification) ya average (regression) leta hai 26 . Isme model training time minimal hota hai (sirf data
store karna), lekin prediction me data se distance compute karni padti hai (slow for large data). K-NN ke liye
distance metric (euclidean, Manhattan, etc.) chunna hota hai.

Hinglish: K-NN me jab koi naya example aata hai, to hum training data se uske sabse kareeb k examples
dhundte hain. Phir un examples ki classes dekhe bina decide karte hain ki naya example kis class ka sabse
jyada nikat hai. Jaise agar 5 najdeek neighbors me 3 “Yes” aur 2 “No” hain, to model “Yes” predict karega.
Distance ke liye hum normal formula use karte hain (jaise Euclidean) 26 .

Udāharan (Hinglish):
- Example: Fruit color aur size ke basis par classifiying fruit: Ek nayi fruit ka size/color dekho, fir do ya
paanch kareeb fruits me dekho ke woh kahan cluster hue the, aur ussi class assign karo.

10
Support Vector Machine (SVM)

English: Support Vector Machine ek powerful supervised algorithm hai jo classification (aur regression) me
use hota hai 27 . SVM ka main goal ek aisa hyperplane (decision boundary) find karna hai jo different
classes ko sabse achhi tarah se separate kare, maximize karta hai margin (hyperplane se nearest points ka
distance) 27 . Support vectors wo training points hain jo hyperplane ke sabse nazdeek hote hain (aur
margin define karte hain) 28 . SVM linear separation ke liye hyperplane (wx+b=0) use karta hai; agar data
linear na ho to kernel trick (RBF, polynomial, etc.) se data ko high-dimensional space me map karke non-
linear boundary bhi draw kar sakta hai. SVM models overfitting se bachne ke liye regularization (C
parameter) use karte hain 27 28 .

Figure: SVM with two classes and maximum-margin hyperplane. Image me do classes (blue vs red) ko separate
karne ke liye purple hyperplane draw kiya gaya hai. Support vectors (closest points) se margin define hota
hai 27 . SVM tries to maximize this margin for best generalization.

Hinglish: SVM me hum ek line (ya hyperplane) nikalte hain jo classes ko separate kare aur is line ke aas-paas
ka gap (margin) sabse bada ho 27 . Wo points jo gap ke paas (border pe) hote hain unko support vectors
kehte hain 28 . In pe focus karke SVM model train hota hai. Agar data linear se separate nahi ho pa raha, to
kernel functions (jaise Gaussian/RBF) se pehle data ko high dimension me map karke phir separate karta
hai.

Udāharan (Hinglish):
- Example: Cat vs dog image classification: Agar ek feature (ear length, tail size) se partially alag ho rahe
hain, SVM apni kernel ke through complex boundary find karega jo dono classes ko achhe se divide karega
27 .

Naive Bayes (Gaussian, Multinomial, Bernoulli)

English: Naive Bayes ek probabilistic classifier hai jo Bayes’ Theorem use karta hai, ye assume karte hue ki
features mutual independent hain 29 . Ye bahut fast aur simple hai, especially high-dimensional text data
ke liye. Teen common variants hain:

• Gaussian Naive Bayes: Jab features continuous hon aur normally (Gaussian) distributed maane
jaen. Gaussian NB har class ke liye feature ka mean/variance estimate karta hai 30 , fir likelihood
calculate karke predictions deta hai.
• Multinomial Naive Bayes: Jab features discrete counts ho (jaise word counts in documents).
Multinomial NB text classification me istemal hota hai, jisme feature represent karte hain ki har word
kitni baar aaya 31 .
• Bernoulli Naive Bayes: Jab features binary (0/1) ho, jaise word-present or not in text documents 32 .
Ye model feature ki presence/absence pe based hota hai.

Har variant me posterior probability P(class|features) Bayes rule se nikalte hain aur sabse zyada probability
wali class predict ki jati hai. Naive Bayes bahut efficient hai, aur kachche text (spam filtering, sentiment
analysis) me best results deta hai 29 .

Hinglish: Naive Bayes ek probabilistic model hai jisme hum kehte hain ki har feature (jaise word frequency)
ek dusre se independent hai 29 . Gaussian NB continuous data ke liye, Multinomial NB word counts ke liye,

11
aur Bernoulli NB binary features (word present/absent) ke liye hota hai 30 33 . Har case me Bayes theorem
se calculate karke predict kiya jata hai.

Udāharan (Hinglish):
- Spam Filtering: Words ki frequency (Multinomial NB) use karke emails ko spam/ham predict karna.
- Document Classification: Document me har term ka count dekh kar topic classify karna (Multinomial NB).
- Binary Features: Web page me kisi keyword ka hona ya na hona ke basis par category predict karna
(Bernoulli NB).

Classification Performance Measures

Confusion Matrix: Binary ya multiclass classification me predictions ko evaluate karne ke liye confusion
matrix use hota hai. Ye ek table hai jisme rows predicted classes aur columns actual classes dikhte hain (ya
vice versa). Binary case me is form ka matrix banta hai:

Actual Positive (P) Actual Negative (N)

Predicted P True Positives (TP) False Positives (FP)

Predicted N False Negatives (FN) True Negatives (TN)

Metrics: In values se kuch important metrics nikalte hain 34 35 36 37 . Kuch pramukh hain: - Accuracy:
(TP+TN)/(TP+TN+FP+FN) – Overall sahi predictions ka proportion 34 .
- Precision: TP/(TP+FP) – Predicted positives me se kitne actual me positive hain 35 .
- Recall (Sensitivity): TP/(TP+FN) – Actual positives me se kitne correctly predict hue 36 .
- F1-Score: 2(PrecisionRecall)/(Precision+Recall) – Precision aur Recall ka harmonic mean 37 . Yeh
imbalanced classes me overall performance batata hai.
- Support: Har class ke actual occurrences ka count (dataset me kitne examples us class ke hain).

Measure Formula Description

Accuracy (TP + TN) / (TP + TN + FP + FN) Overall correctness of model 34

Correct positives among all predicted positives

Precision TP / (TP + FP)
35

Recall Correctly identified positives among all actual

TP / (TP + FN)
(TPR) positives 36

2 × (Precision * Recall) / (Precision +

F1-Score Harmonic mean of precision and recall 37
Recall)

Hinglish: Confusion matrix me dekhte hain ki model ne kitne Positive sahi-buray aur Negative sahi-buray
predict kiye. Phir accuracy (kul sahi predictions ka ratio) 34 , precision (predicted positive me se sahi kitne)
35 , recall (actual positives me se sahi kitne) 36 , aur F1-score (precision/recall ka harmonic mean) calculate

karte hain 37 .

Udāharan (Hinglish):
- Agar 100 patients me se model ne 90 sahi predict kiye, to accuracy 90% hogi.

12
- Agar model ne 20 patients ko disease bataya aur 15 me sach tha, to precision 15/20=75% 35 .
- Agar total 30 patients me disease tha aur model ne 15 sahi bataya, to recall 15/30=50% 36 .
- F1 score dono ko combine karke ek single metric deta hai.

Unit IV: Ensemble Learning and Random Forests

Ensemble Learning me hum ek strong model banane ke liye bahut saare weaker models ko combine karte
hain. Jaise hamari life me ek se zyada logo se advice lena zyada accha decision de sakta hai, waise hi ML me
multiple models ka answer zyada accurate hota hai 38 . Ensemble methods se predictive accuracy aur
robustness improve hoti hai.

Hinglish: Ensemble me hum “committee” approach lete hain — kai models train karke unki outputs combine
karte hain 38 . Individual model galti kar sakta hai, par group milke usually behtar faisla deta hai.

Udāharan (Hinglish):
- Example: Teen alag decision trees banakar unke votes se final class decide karna (voting ensemble). Har
tree alag data subset se banega, milake consensus dega.

Basic Ensemble Techniques

• Voting and Averaging: Sabse seedha ensemble technique hai. Classification me majority voting
hota hai: alag-alag classifiers ki predictions lete hain, aur sabse zyada aaye huye class ko final output
maante hain 39 . Regression me averaging hota hai: models ke predictions ka average ya weighted
average nikalte hain. Is se variance reduce hoti hai aur accuracy badhti hai 40 41 .

• Max Voting / Weighted Voting: Kuch ensembles me sab models equal weight nahi hotey. Aap unko
performance ke hisaab se weight de sakte hain aur weighted voting kar sakte hain 41 . Matlab accha
perform karne wale model ko zyada awaz milti hai final decision me.

• Bagging (Bootstrap Aggregating): Bagging me hum dataset ke multiple bootstrap samples

(random subsets with replacement) banate hain. Har subset par separate model train hota hai (often
decision trees). Fir un models ke outputs ko combine karte hain (majority vote for classification,
averaging for regression) 40 39 . Bagging se variance decrease hota hai (overfitting kam hota hai)
40 . Random Forest (neeche) ek khas bagging ensemble hai.

• Pasting: Bahut similar to bagging, bas replacement ke bina random subsets banata hai. Training
subsets me thora antar hai, par concept same hai.

Hinglish: Voting me har model apna vote deta hai, aur sabse zyada votes wali class chun lete hain. Bagging
me original data me se kai random subsets lete hain aur har subset pe ek model train karte hain 40 ; phir
sabka vote ya average lete hain.

Out-of-Bag (OOB) Evaluation

Bagging me jab bootstrap sampling se subsets bante hain, kuch samples har model ke training me use nahi
hote (unko out-of-bag data kehte hain). In OOB samples ka istemal karke model ki performance estimate ki

13
ja sakti hai bina extra validation set ke 42 . Matlab har model ke liye jo data nahi pada, us par test karke
overall accuracy gauge karte hain.

Hinglish: OOB samples vo data hote hain jo ek particular bagging subset me nahin aaye. Unhein model ki
test set maan ke accuracy measure karte hain 42 . Ye ek tarika hai bina alag data ke performance dekhne
ka.

Random Patches and Subspaces

• Random Subspace Method: Har model training me features ki random subset use karta hai, na ki
saare features 43 . Isse models diverse bante hain kyunki har model alag features se train hota hai.
Feature-level randomness se generalization aur overfitting control hota hai.

• Random Patches: Ye combined approach hai — random instances (bootstrapping) ke saath-saath

features ki random subsets bhi. Is se variation badh jata hai models me, further error reduce hota
hai.

Hinglish: Random subspace me har classifier sirf kuch features dekhta hai 43 . Jaise ek tree sirf age aur
income dekh ke, doosra age aur loan history dekh ke. Sab alag nazariye se train honge. Random patches
me instance aur feature dono random hote hain, jisse models aur bhi different ban jate hain.

Random Forests (and Extra-Trees, Feature Importance)

Random Forest: Ye bagging ka ek special case hai jisme base learners decision trees hote hain. Random
Forest me har tree ko bootstrap sample aur random subset of features (subspace) dono milte hain. Har tree
independent train hota hai, aur final prediction me sabke votes/averages liye jate hain. Ye overfitting kam
karta hai aur variance bahut reduce karta hai 44 . GfG kehta hai: Random Forest multiple decision trees
banata hai aur unke predictions ko aggregate karta hai, jo final output deta hai 44 .

Extra-Trees: Ye Random Forest ka variant hai jisme trees banate waqt randomness aur badh jati hai (har
split ke liye thresholds randomly choose kiye jate hain). Isse speed tez hoti hai aur bhi variance kam hota
hai.

Feature Importance: Random Forests me hum dekh sakte hain ki kaunse features model ke liye sabse
zyada maayne rakhte hain (kaunse frequently tree splits me use huye). Ye feature importance measure
karke batata hai ki prediction me konsa feature kitna contribute karta hai.

Hinglish: Random Forest me decision trees ko bootstrap samples aur feature subsets diya jata hai 44 , phir
sabke answers vote/average se final aata hai. Har tree thoda alag hota hai, isliye overall result zabardast
hota hai. Extra-Trees me trees aur random bante hain (even splits random). Forest se hum feature
importance bhi nikal sakte hain — jo features splits me zyada aate hain, wo important hote hain.

Udāharan (Hinglish):
- Agar insurance risk model banana ho, to Random Forest alag-alag subsets of customers aur features (age,
driving score, car model, etc.) pe trees banata hai, fir un sabke votes se “High risk/Low risk” decide karta hai.
Ye kaafi robust approach hai.

14
Boosting (AdaBoost, Gradient Boosting)

Boosting: Isme models ek ke baad ek sequentially bante hain. Har naya model un examples par focus karta
hai jahan pehle wale models galat the. Models ki predictions weighted tarike se combine hoti hain 45 .
Boosting se bias reduce hota hai (complexity badh jati hai) aur overall error kam hota hai.

• AdaBoost (Adaptive Boosting): Yahan har data point ko ek weight milta hai. Pehla weak classifier
banate hain, jisme jo points galat predict hue unke weights badha diye jaate hain. Fir second
classifier un high-weight points pe zyada focus karta hai. Final prediction weak classifiers ke
weighted vote se aata hai 46 .

• Gradient Boosting: Ye trees ko sequentially add karta hai jahan har naya tree previous trees ke
errors ko correct karta hai 47 . Ye loss function (e.g. squared error, log-loss) ke gradient ke opposite
direction me train hota hai. Gradient Boosting machines (GBM) bohot powerful hai (XGBoost,
LightGBM ye advanced versions hain).

Hinglish: Boosting me hum ek ke baad ek tree/trainers banate hain. AdaBoost me pehle wale errors ko dekh
ke agla model unpe dhyan deta hai 46 ; Gradient Boosting me har tree previous tree ke mistakes ko kam
karta hai 47 . Final prediction in sab models ke weighted sum hota hai. Ye overfitting kam karne aur
accuracy badhane me kaafi effective techniques hain.

Udāharan (Hinglish):
- Example (AdaBoost): Pehle ek simple tree banao (threshold ka use karke), uski galtiyon ko identify karo,
phir doosra tree banakar un galtiyon ko sudharo 46 . Aise teen-char trees milke ek strong classifier banate
hain.

Stacking (Stacked Generalization)

English: Stacking me multiple different models (base-learners) train karte hain, phir unki predictions ko ek
meta-learner (higher-level model) ke input bana ke final output nikalte hain 48 . Meta-learner (jaise logistic
regression ya decision tree) seekhta hai kaunse base model ke predictions ko kaise weight karna hai.
Stacking ensemble often bagging/boosting se bhi behtar results deta hai, magar careful cross-validation se
meta-model train karna hota hai.

Hinglish: Stacking me sab alag type ke models train karte hain (decision tree, SVM, neural net, etc.), fir unki
predictions ko gather karke ek final model ko aur train karte hain. Matlab base models ka ensemble build
hota hai jo fir meta model ko input deta hai, taaki final decision aur accurate ho 48 .

Udāharan (Hinglish):
- Example: Teen models: logistic regression, SVM, aur k-NN train karo. Un sab ke predicted probabilities ko
dataset me add columns banake meta model (jaise decision tree) train karo jo final answer dega.

15
Unit V: Dimensionality Reduction and Learning Theory

The Curse of Dimensionality

English: Curse of Dimensionality se matlab hai high-dimensional data me kaafi challenges aana 49 . Jaise
jaise features (dimensions) badhte hain, data sparse ho jata hai aur patterns dhundhna mushkil hota hai.
Computational complexity aur training time exponential badh jate hain, aur overfitting ka risk bhi. GfG
batata hai ki high dimensions me data points ka distance uniformly barh jata hai, jisse learning model
confuse ho sakta hai 49 . To handle isliye dimensionality reduction ya feature selection strategies zaroori
hain.

Hinglish: Jab features bohot saare hain (jaise 1000), to data bahut sparse ho jaata hai 49 . Machine Learning
me phasaad hota hai — training slow ho jaati hai, model ko jyada examples chahiye, aur accidents bhi badh
sakte hain. Is problem ko “curse of dimensionality” kehte hain.

Udāharan (Hinglish):
- Example: Ek image with 1024 pixels me classifier train karne se pehle chahe bohot data ho, patterns
samajhna mushkil hota hai. Ek hi cluster me points bahut door door milte hain, learning degrade ho sakti
hai.

Approaches: Projection and Manifold Learning

Projection Methods: Linear techniques jo high-dimensional data ko lower-dimensional subspace me

project karte hain (jaise PCA). Ye assumes karte hain ki data ka major variation kisi linear subspace me hai,
isliye linearly project karke information preserve ki ja sakti hai.

Manifold Learning: Non-linear techniques assume karte hain ki data kuch lower-dimensional manifold
(curve/surface) par lie karta hai within high-dimensional space. Algorithms jaise t-SNE, Isomap, UMAP, LLE
in manifolds ko find karne ki koshish karte hain.

Hinglish: Projection me hum data ko ek seedhi line (ya plane) me compress karte hain, Manifold Learning
me data ko curved surface pe compress karte hain.

Udāharan (Hinglish):
- Faces ka data: Agar har face ko 100×100 pixel image se 10-sahi values (principal components) me
compress karna hai, to PCA use hoga. Agar non-linear shapes (jaise oscillating curve data) detect karna ho,
to t-SNE jaisa method better hoga.

Principal Component Analysis (PCA)

English: PCA ek popular projection technique hai jisme hum data ke principal axes dhundte hain jo
maximum variance capture karte hain 50 . Ye method data ka covariance matrix compute karke uske top
eigenvectors (principal components) leta hai. Pehla principal component data ko sabse zyada spread

16
(variance) ke direction me rotate karta hai; dusra component bacha hua maximum variance capture karta
hai, etc 50 .

• Preserving Variance: PCA ka goal hota hai ki reduced dimensions me bhi jyada information
(variance) preserve ho. Hum eigenvalues dekh ke explained variance ratio nikal sakte hain (kaunsa
PC kitna variance explain karta hai) 50 .
• Projecting Down: Data ko first d principal components ki directions pe project karke d-dimension
me laate hain. Agar cumulative explained variance (e.g. 95%) threshold rakha ho, to utni PCs choose
karte hain ki itna variance cover ho jaye.
• Choosing Dimensions: Scree plot (eigenvalues ka plot) ya explained variance plot se decide karte
hain ki kaunse number of components rakhna hai.
• Compression: PCA se data ka compression bhi hota hai — large dataset ko fewer dimensions me
store karke memory bachti hai, aur noise bhi kam ho sakta hai.
• Randomized/Incremental PCA: Bade datasets ke liye variations hain (like Randomized PCA using
random projections for speed, Incremental PCA that processes chunks of data) takki memory aur
time efficient ho jaye.

Hinglish: PCA ek linear method hai. Pehle principal component sabse zyada variance capture karta hai;
dusra component baaki variance ko. Data ko in principal axes pe project karke hum uska dimension reduce
kar lete hain. Example: Agar image dataset ke 1000 features hain, PCA se k components me reduce karke
bhi 95% information rakh sakte hain 50 .

Udāharan (Hinglish):
- Example: Face recognition me hum har face (100×100 pixel) ko PCA ke first 50 components me compress
kar dete hain; tab bhi faces ko pehchanna possible hota hai, aur model train faster hota hai.

Kernel PCA

English: Kernel PCA PCA ka non-linear extension hai. Isme data ko pehle ek kernel function (jaise
Gaussian/RBF, polynomial) se higher-dimensional feature space me map karke linear PCA apply karte hain.
Kernel PCA complex (non-linear) patterns ko capture kar sakta hai jinko normal PCA handle nahi kar sakta.
Kernel aur uske hyperparameters (jaise RBF ka gamma) chunna padta hai, jo cross-validation ya prior
domain knowledge se decide kiye ja sakte hain.

Hinglish: Kernel PCA me hum pehle data ko ek kernel (RBF ya polynomial) se map kar dete hain, phir PCA
apply karte hain. Isse aise non-linear structures bhi capture hoti hain jinko plain PCA nahin nikal sakta.

Udāharan (Hinglish):
- Example: Agar ek spiral-shaped data hai, normal PCA fail karega; lekin kernel PCA (RBF) se aap spiral ko
linear space me unfold karke compress kar sakte hain.

Learning Theory: PAC and VC Model

PAC Learning (Probably Approximately Correct): Ye framework batata hai ki hum ek class of functions ko
kaise efficiently learn kar sakte hain. Algorithm PAC-ly learn karega agar polynomial number of samples se,
confidence (1−δ) ke saath error ≤ ε achieve kar sake. Informally, “with high probability, the learned
hypothesis is almost correct”.

17
VC Dimension: VC (Vapnik–Chervonenkis) dimension ek measure hai model ki capacity/complexity ki. Ye wo
maximum number of points ka size hai jo model class shatter (har possible labeling achieve) kar sakta hai
51 . High VC dimension matlab model bahut flexible hai (complex boundaries), low VC dimension matlab

simple model. For example, linear classifier (line) ka VC dimension typically 3 (3 points in general position
can be shattered in 2D), whereas a high-degree polynomial classifier ki VC dimension zyada hoti hai 51 .

Hinglish: PAC learning ke mutabik hum chaahte hain ki hamara learner jitna ho sake sahi ho (ε error tak) aur
jitna ho sake sure ho (1−δ confidence) given enough data. VC dimension ek theory measure hai jo batata hai
ki model kitna complicated hai. Agar VC dimension bada hai, model training data ko bahut alag tareeke se
classify kar sakta hai 51 , lekin usse zyada data ki bhi zaroorat hai. Simple models jaise linear functions ka
VC dimension chhota hota hai, complex models (bohot flexible) ka VC dimension bada.

Udāharan (Hinglish):
- VC Example: Agar ham sirf linear separators (lines) use kar rahe hain, to teen points ko har tarah label kar
sakte hain shatter karke 51 , magar chaar points nahi (ek arrangement me).

1 13An Intuitive Explanation of Data Science Concepts. Part I: DS, AI, ML, and DL in Simple Words ■ Data
Science UA
[Link]

2 4 What is Machine Learning? Definition, Types, Tools & More | DataCamp

[Link]

3 Supervised, Unsupervised, and Semi-supervised Learning: What’s Right for Me? - ECS Equus Compute
Solutions
[Link]

5 6 12 Bias-Variance Trade Off - Machine Learning - GeeksforGeeks

[Link]

7 Why Data Visualization is Essential in Every Step of ML

[Link]

8 Hypothesis in Machine Learning

[Link]

9 10 What is data augmentation? | IBM

[Link]

11 Data Normalization Machine Learning - GeeksforGeeks

[Link]

14 15 16 18 19 Clustering in Machine Learning - GeeksforGeeks

[Link]

17 22 What is a Gaussian Mixture Model? | IBM

[Link]

20 ML | BIRCH Clustering - GeeksforGeeks

[Link]

18
21 Basic Understanding of CURE Algorithm - GeeksforGeeks
[Link]

23 Maximum a posteriori estimation - Wikipedia

[Link]

24 Logistic Regression in Machine Learning - GeeksforGeeks

[Link]

25 Decision Tree in Machine Learning - GeeksforGeeks

[Link]

26 K-Nearest Neighbor(KNN) Algorithm - GeeksforGeeks

[Link]

27 28 Support Vector Machine (SVM) Algorithm - GeeksforGeeks

[Link]

29 30 31 32 33 Naive Bayes Classifiers - GeeksforGeeks

[Link]

34 35 36 37 What is Accuracy, Precision, Recall and F1 Score?

[Link]

38 39 40 41 42 43 44 45 46 47 48 Ensemble Learning - GeeksforGeeks

[Link]

49 Curse of Dimensionality in Machine Learning - GeeksforGeeks

[Link]

50 Principal component analysis - Wikipedia

[Link]

51 Vapnik–Chervonenkis dimension - Wikipedia

[Link]

19