Lecture 3 - Multi-class classification and regression

Question 1

Give an example of multi-class classification

Answer 1

1. Disease type diagnosis
2. topic classification

Question 2

what are the two approaches to turn binary classifier into multi-class

Answer 2

1. One versus rest
2. One versus one

Question 3

What is one versus rest

Answer 3

Each classifier distinguishes between one specific class and all other classes combined. The class with the highest confidence score from its respective classifier is chosen as the final prediction.

Question 4

what is the process of learning and inference one-versus-rest

Answer 4

Learning: traink or k-1 seperate classifiers where k is the number of classes
Inference: use all the and form a code word based on the output of the classidier. Next compare the code word against all the rows and dinf the cnearest row in the code matrix.

Question 5

what is one-versus-one

Answer 5

strategy for multi-class classification in machine learning where a separate binary classifier is trained for every possible pair of classes.

Question 6

For n classes, this results is n(n−1)/2 classifiers for symmetric, and n(n-1) for asymetric.

who does this apply to?

Answer 6

one versus one

Question 7

what is the process of training and inference of one-versus-one

Answer 7

Training: traiin seperate classifiers for each pair of classes
Inference: use all the classifications for a code word based on the output of the classifier. Next, compare the code word against all rows and find the nearest row in the code matrix. Take a voting scheme when distances are not unique.

Question 8

How to get the accuracy in a confuzion matrix of a three-class confuzion matrix

Answer 8

add up all the True Positives divided by the total

Question 9

How to get the precision in a confuzion matrix of a three-class confuzion matrix

Answer 9

get the precision for each class, then multiply it by the distribution with the total, then add all the three weighted precisions together.

Question 10

how can we know how good a classifier can be?

Answer 10

1. macro-average
2. micro-average

Question 11

what is macro-average

Answer 11

macro-averate will compute the metric independently for each class and then take the average.

Question 12

what is micro-average

Answer 12

micro-average will aggregate the contributions of all classes to compute the average metric.

Question 13

how should AUC curves be used for multi-class classifiers

Answer 13

The average AUC over binary classification tasks, eigher in a one-versus-rest of one-versus-one.

Question 14

what does ROC stand for

Answer 14

Receiver operating characteristics

Question 15

What is regression loss

Answer 15

Regression models are evaluated by applying a loss function to the residuals. f(x)- ^f(x)

Question 16

how many parameters does n-degree polynomial have?

Answer 16

n+1 parameters

Question 17

how to avoid overfitting in regression

Answer 17

To avoid overfitting, the number of parameters estimated from the data must be considerably less than the number of data points.

Question 18

what is the Bias-variance dilemma

Answer 18

A low-complecity models suffers less from variability due to random variations in the training data, but mey introduce a systematic bias that even large amounts of training data can’t resolve; on the other hand, a high-complexity model eliminates such bias but can suffer non-systematic errors due to variance.

Question 19

In ____ learning the task if to come up with a description of the data

Answer 19

descriptive

Question 20

what is distance based clustering

Answer 20

Most distance-based clustering methods depend on the possibility of defining a ‘centre of mass’ or exemplar of an arbitrary set of instances, such that the exemplar minimises some distance-related quantity over all instances in the set, called its scatter. A good clustering is then one whereby the scatter is summed over each cluster - the witithin-cluster scatter is much smaller than the scatter of the entire data set.

Question 21

What is purity

Answer 21

check slides or gpt
1/N sigma(max(omega_k union c_j))

Question 22

What are the three ways to evaluate clustering performance without ground truth?

Answer 22

1. Calinski-Harabsz Index
2. Davies-Bouldin Index
3. Silhouette Coefficient

Question 23

what is the silhouette coef formula

Answer 23

s=(b-a)/max(a,b)

a: the mean distance between an instance and all other points in the same cluster
b: the mean distance between an instance and all other points in the next cluster.

Question 24

what is the range of silhouette coef results

Answer 24

-1 to +1. +1 far from neighbours, 0 on the decision line, -1might be assignmed to the wrong cluster.

Question 25

Give 2 examples of subgroup discovery

Answer 25

1. Detection of risk groups with coronary heart disease or cancer
2. Finding patterns in traffic accidents

Question 26

how can we assess performance of sub-group discovery

Answer 26

1. Chi squared test
2. class distribution on the left different from class distribution in the row marginals