week 1

Question 1

prosecutor’s fallacy

Answer 1

P(A|B) != P(B|A)

Question 2

data dredging

Answer 2

performing many statistical results and only reporting those with significant results / claiming that a model learned from data represents the truth

Question 3

bonferroni’s principle

Answer 3

even in completely random datasets, you can expect particular events of interest to occur, and to occur in increasing numbers as the amount of data grows

Question 4

power laws

Answer 4

linear relationship between the logs of 2 variables, a lot of real-world data follows this rule. to lessen the effect of very large values, NEED TO log

Question 5

data transformation

Answer 5

box-cox transform - family of power functions - result depends on lambda

Question 6

feature extraction

Answer 6

embedding

Question 7

feature transformation

Answer 7

categorical : label encoding, target encoding (sort by class probability), one-hot encoding(1 column for each class, binary can also be possible)
continuous : equal range, equal frequency, clustering, classification, reduce dimensionality and cluster latent space

Question 8

feature addition

Answer 8

x^2 or x(t) - x(t - 1), but be careful to NEVER USE CLASS LABELS FROM TEST SET

Question 9

data sampling

Answer 9

99% of data is benign, 1% is fraudulent => 99% accuracy => data sampling
techniques : oversample minority class, undersample majority class, reweighting, synthesizing
IMPORTANT : NEVER MODIFY TEST SET

Question 10

SMOTE

Answer 10

for every minority instance, take a random neighbour, construct vector, choose new random artificial data point along that vector. before this algorithm you can remove tomek links (links between nearest neighbours of different classes)

Question 11

missing values

Answer 11

imputation (replacing missing values) :
hot-deck - replace with values from similar rows
multiple imputing