Machine Learning

Question 1

What does ML allows us to do

Answer 1

By using data from past observations we can predict future outcomes or values
ex. icecream sales based on historical sales and weather records

Question 2

What is function in ML terms

Answer 2

ML encapsulates a function to calculate output value based on one or more input values

process of defining function is known as training
predict new values based on function is known as inferencing

Question 3

ML Steps for Training and Inferencing

Answer 3

1. Obtain training data of past observations. This data has the attributes or features(x) of thing being observed and known output (label(y)). [x1,x2,x3], y … lots of this data

ex. for icecream sales case features could be temp, rainfall, windspeed and # sold would be the label

2. Algorithm is applied to data to determine relation between features and label. Specific algorithm applied depends on the problem, but main goal is to fit a function to the data
3. result of algorithm is a function that captures the model. y= f(x). We can use this to make predictions now
4. Training phase is complete. Now we can perform inferencing. Use our function to get predicted value represented by y hat (y^)

Question 4

What are the different types of Supervised ML?

Answer 4

Regression and Classification (Binary and Multiclass)

Question 5

What is an example of Unsupervised ML?

Answer 5

Clustering

Question 6

Supervised ML

Answer 6

Supervised ML is when the training data includes both feature values and labels

Question 7

Regression

Answer 7

label predicted by model is a numeric value

i.e ice cream sales, fuel efficiency, property price

Question 8

Binary Classification

Answer 8

label determines whether observed item is or isn’t something.

i.e. is patient diabetic based on weight, age, blood
i.e.2. will customer default on loan based on income, age, credit history

Question 9

Multiclass Classification

Answer 9

Extension of binary classification. to determine multiple outcomes.

i.e out of 3 species which penguin can this be based on physical measurements
genre of movie based on cast, director, and budget

Question 10

Unsupervised ML

Answer 10

involves training models that consist only of feature values without any known labels. The ML algorithms will determine relation between features to group them

Question 11

Clustering

Answer 11

most common form of unsupervised ML.
clustering algorithm identifies similarities between observations based on their features and groups them into discrete clusters

i.e. group similar flowers based on size, number of leaves and number of petals
It can be thought of to be similar to multiclass except we do not have the defined labels with clustering

Question 12

Regression Model process

Answer 12

1. split training data to data you will train with and subset to validate trained model
2. use algorithm to fit training data to model - use regression algorithm like linear regression
3. use validation data held back to test the model by predicted labels for features
4. Compares actual labels for validation data against predicted. Then aggregate differences between predicted and actual label values to see how the model performs

This process can be repeated with different algorithms and parameters.

Question 13

What are some evaluation metrics for regression?

Answer 13

Mean Absolute Error (MAE)
Mean Squared Error (MSE)
Root Mean Squared Error (RMSE)
Coefficient of Determination (R^2)

Question 14

What is calculated in classification model

Answer 14

probability values — instead of numeric values like in regression
for binary classification the value is 1 or 0 (yes or no, true/false)

Question 15

What kind of shape does the function for binary classification take?

Answer 15

it takes a sigmoid shape kind of like an S and one algorithm that can be used to get this is logistic regression (top and bottom part do look like a logarithmic function)

Question 16

How is binary classification evaluated

Answer 16

using a matrix, we can see how many items were in correct and incorrect positions
true negatives, false positives, false neg, true pos. also called confusion matrix.

using confusion matrix we can do math for:
Accuracy: tn+tp/(tn+fn+fp+tp)
Recall: tp/(tp+fn)
Precision: tp/(tp+fp)
F1-Score: 2PrecisionRecall/(Precision + Recall)

Question 17

Are processes different for multiclass classficiation?

Answer 17

No, follows the same as binary and regression with iterative train, validate and evaluate process

Question 18

How do multiclass algorithms work?

Answer 18

calculate probability for multiple class labels

Question 19

What type of multi-class algorithms can you use

Answer 19

one-vs-rest (OvR) algorithm
Multinomial algorithm

Question 20

How does OvR algorithm work

Answer 20

We train a binary classification function for each class, that calculates probability of it being that class vs another class. Each algorithm produces a sigmoid function wiht values between 0 and 1. The model using this algorithm would select the class that produces highest probability output.

Question 21

How does multinomial algorithm work?

Answer 21

creates a single function that returns multi-valued output. Output is a vector (array of numbers) [0.2, 0.3, 0.5] that all add up to 1. ex function is softmax f’n.

We also use confusion matrix 3x3 for 3 classes example.

Question 22

Deep Learning

Answer 22

advanced form of ML that tries to copy human brain learning. creates artificial neural network