Python ML Principles

Question 1

What are the four main steps in ML?

Answer 1

Visualisation
Cleaning and Transformation
Construction of ML model
Evaluation of ML model

Question 2

What are the two main sub-steps of the Cleaning and Transformation step?

Answer 2

1) Data Preparation & Cleaning

2) Feature Engineering

Question 3

What should you do before starting Preparation & Cleaning?

Answer 3

Explore the data to understand the issues that are present.

Question 4

What are six sub- steps of the Data Preparation & Cleaning step?

Answer 4

1. Recode chr strings to eliminate unrecognised characters
2. Find & treat missing values
3. Set correct data type and column
4. Transform categorical features to increase cases
5. Apply transformation to numerics to improve distributions
6. Duplicate management

Question 5

What’s another name for “transformation to improve distribution” ?

Answer 5

Feature engineering

Question 6

Name a common transformation.

Answer 6

Log

Question 7

What is the main thing we’re trying to achieve with Feature Engineering?

Answer 7

We’re trying to achieve distinct separation of the labelled cases, indicating better prediction.

Question 8

What is a test used to evaluate linear regression accuracy?

Answer 8

Sum of squared errors.

Question 9

Why is linear regression sometimes called Least Squares Regression?

Answer 9

Because it creates a line that minimises the square of variance (error) from the line.

Question 10

With linear regression in scikit.learn what Python package should you use for your arrays?

Answer 10

Numpy

Question 11

What are the 4 main steps for linear regression with scikit.learn?

Answer 11

Layout numpy arrays
Scale
Specify model object
Fit

Question 12

What visualisation could you use to evaluate residuals of a regression model?

Answer 12

A histogram of residuals.

Question 13

What two residuals histogram pattern indicate an accurate model?

Answer 13

1. Clustering of residuals around zero.

2. Normal distribution.

Question 14

If you see a multimodal residuals histogram what should you do about the non-zero modes?

Answer 14

Investigate what’s creating them and consider adding these features to your model.

Question 15

In scikit learn, what is onehot encoding?

Answer 15

Conversion of multiple feature options to a numpy array, where only one row for the record shows a “1”

Question 16

In classification ML, what is the space between the two outcomes called?

Answer 16

The decision boundary.

Should be zero, with Y/N result giving -ve +ve values.

Question 17

What is a loss function?

Answer 17

A weighting line in classification that describes how much incorrect labels are to be penalised.

Question 18

Name four ways to check if your model is CRAP.

Answer 18

Confusion matrix (TP,FP,TN,FN)
ROC curves
Accuracy/misclassification error
Precision/Recall/F1

Question 19

What is on the y-axis and what on the x-axis of an ROC curve?

Answer 19

y-axis = TPR
x-axis = FPR

Question 20

Name 4 techniques to address imbalanced data.

Answer 20

1) Undersample majority
2) Oversample minority
3) Case weights
4) Impute

Question 21

Name one method of dimensionality reduction.

Answer 21

Principal component analysis.

Question 22

What data does Principal Component Analysis produce on the dataset features?

Answer 22

Variability in the label explained by each feature.

Question 23

What is the goal of regularisation?

Answer 23

Prevent overfitting of ML models.

Question 24

What is the tradeoff dilemma in Regularisation?

Answer 24

Bias vs Variance

Question 25

Regularisation reduces variance, but can introduce what?

Answer 25

Bias

Question 26

A model that has high variance and low bias is what?

Answer 26

Overfit

Question 27

The diagonal line on a QQ Residual vs Predicted plot represents what?

Answer 27

A perfect Normal Distribution.

Question 28

What’s another name for L2 regularisation?

Answer 28

Ridge Regression

Question 29

What does L2 regularisation do to the coefficients?

Answer 29

Constrains them, by driving coefficients close to zero.

Question 30

What are two other names for L1 regularisation?

Answer 30

Lasso method.

Manhattan norm.

Question 31

What is k-fold cross-validation?

Answer 31

Resampling random subsets and repeating training calculations, while holding 1 subset back for testing.

Question 32

With decision trees, what kind of split creates less entropy?

Answer 32

A split where there are very uneven probabilities of the outcomes.

Entropy is high = 1 if p = 0.5.