Algorithms

Question 1

Perceptron

Answer 1

• algorithm
• weight update (1 - 0 ,0 - 1)
• graphical representation
• threshold function

Question 2

Sigmoid Function

Answer 2

• parametric, discriminative
• formula = 1/(1+e^(-z))
• 0 <= sigma(z) <= 1

Question 3

Hold out method

Answer 3

• split data into train test

- do this k times

Question 4

k fold cross validation

Answer 4

• n data points
• split into k folds(n/k)
• each with equal size
• each fold must be balanced
• use k-1 fold as train and 1 as test
• repeat k times

Question 5

leave one out method

Answer 5

• n data points
• n-1 points in train
• 1 as test
• repeat n times

Question 6

Performance Measure

Answer 6

Accuracy = (tp+tn) /(p+n)
Precision (P)
    (+) =  tp / pp
    (-) = tn / pn
Recall (R)
    (+) = tp / p
    (-) = tn / n
F measure = 2PR / (P+R)
TPR = tp / p
FPR = fp / n 
ROC 
AUC = h(a + b)/2 

tp - true positive
tn - true negative
p - positive in dataset
n - negative in dataset
pp - predicted positive
pn - predicted negative

Question 7

Probabilistic Classifier

Answer 7

NB, Logistic Regression, Sigmoid

Question 8

Discrete Classifier

Answer 8

Perceptron

Question 9

KNN

Answer 9

• classification, regression
• supervised learning
• non parametric
• no model at train phase
• closeness/ proximity
• algorithm
• lazy learner
• requires normalization
• instance based learning
• sensitive to noise and outliers
• larger k produces smooth boundaries

Question 10

NB

Answer 10

• generative model
• probabilistic classifier
• discrete -based on conditional probabilities
• continuous -pmf, pdf
• zero frequency problem
• smoothing (add one smoothing, Laplacian smoothing)
• bivariate and multivariate ND
• Mahalanobis distance

Question 11

Decision tree

Answer 11

• vertices are features