Machine Learning

Front 1

How to obtain the probability of getting two specific outcomes if the events are mutually independent?

Back 1

Multiply the two probabilities

Front 2

Bayes Equation for P(x|Oc)?

Back 2

P(x|Oc)= P(Oc|x) * P(x) / P(Oc)

Front 3

In P(x|Oc) , how is it P(Oc) called?

Back 3

Marginal likelihood

Front 4

What is Marginal Likelihood and why it is not so important?

Back 4

In P(x|Oc) the Marginal Likelihood is P(Oc). When comparing two posterior probabilities for the same problem it is not that important because, both will have the same Marginal likelihood, so the factor between them will be the same (P1 will still be 10 times grater than P2). If you are want exact values rather than comparing then it is important.

Front 5

What is the likelyhood?

Back 5

Its a probability than can be derived from data alone. It doesn't take into account previous knowledge.

Front 6

What is the posterior probability?

Back 6

Posteriors probability depends not only data but in previous experience.

Front 7

What does PCA does?

Back 7

Founds components (vectors) in the direction that maximizes the variance of the data sample compared to itself (not to a bigger group). PCA finds several vectors that are perpendicular to each other.

Front 8

What is normalize?

Back 8

Rescale all values of a sample between 0 and 1.

Front 9

What is Standarization?

Back 9

Rescale all values by subtracting the mean and dividing by the std deviation. It will leave the mean of the population in 0 and the standard deviation in 1.

Front 10

What t-SNE?

Back 10

It is a graphical representation technique in 2D of an N-dimension dataset. It can give you some information about how close the samples are, even could visually separate samples in clusters.

Front 11

What values of Learning rate to use to test t-SNE?

Back 11

between 50 and 200.

Front 12

What is the null hypothesis of Pearson Correlation?

Back 12

There is no statistically significant relationship between the [insert variable 1] and [insert variable 2].

In other words: Variable 1 varies independently from Variable 2

Front 13

What does a low PValue means?

Back 13

That the null hypothesis is wrong.

Front 14

What is a confusion Matrix?

Back 14

Front 15

Formula of precision and what does it means?

Back 15

High precision means that when you say something is of the being evaluated, you are right most of the time. Is how good is your possitive prediction with respect to all your possitive predictions.

Front 16

Formula Recall?

Back 16

High Recall means that most span emails are classified correctly

Front 17

Formula F1 score?

Back 17

Front 18

What are the 4 C's of Machine Learning?

Back 18

Correcting (fixing wrong values like 800 in age it is probably 80, or N in sex is probably male), Completing (fillna with mean for example), Creating (creating features from others, bins with qcut and cut), Converting (formating, string to categorical)

Front 19

What does pd.qcut does?

Back 19

Divide the range of values in a sample (from min to max) in X ranges. All ranges will have the same amount of samples but they their length can vary. In other words it founds the limits of the quantiles. If X = 4, it founds the 4 value ranges which are the boundaries of the quartiles. Then it maps each value to the range it is in.

Front 20

What does pd.cut does?

Back 20

It divides the sample range (from min to max) in find X (a param) ranges of the same length (max-min = same for all ranges) and maps each sample to the range it is in.

Front 21

What is a quantile?

Back 21

It is a division of the sample in a way that all divisions has the same number of sample. If the sample is divided in 4 quantiles it is called a quartile.

Front 22

How to calculate Inter Quartile Range?

Back 22

IQR = Q3 − Q1

Front 23

What is a confusion matrix?

Back 23

See picture. Reference is the actual response (y values) and prediction what you predicted

Front 24

What is a good tool to see which features are good to cluster or classify?

Back 24

The density plot. It shows you how much each feature overlaps in each class. Features that don't overlap much, e.g. Petal.length and Petal.width in the picture, are good for classification and clustering.

Front 25

When using train/validation/test, what temptation you need to resist when comparing models?

Back 25

If we have 2 models (doesn't matter if they use different algorithms or is the same with different hyper params values), and 1 performs best with validation data, that is the one you want to use. You don't want to fall in the temptation of choosing the one that perform best in test data.

Front 26

What package you can use in R for kfold and tunning for classification and regression?

Back 26

caret package has a train method

knn_fit <- train(as.factor(V11)~V1+V2+V3+V4+V5+V6+V7+V8+V9+V10, data, method = "knn", # choose knn model trControl=trainControl( method="repeatedcv", # k-fold cross validation number=10, # number of folds (k in cross validation) repeats=5), # number of times to repeat k-fold cross validation preProcess = c("center", "scale"), # standardize the data tuneLength = kmax) # max number of neighbors (k in nearest neighbor)