Multi-Label Classification

Question 1

What is the basic idea of multi-class labels?

Answer 1

Each example can be assigned one or more class labels

Question 2

Example: Class label = Different types of disease in hospital patients

Answer 2

So what might be covered in a table, for example, is Cancer, and Diabetes, both are class labels. Look out for this in the exam (you will have multiple class labels that someone wants to find out)

Question 3

What are the two approaches for Multi-Label Classification?

Answer 3

Problem transformation approach (this one will be focused on)

Algorithm-adaptation approach

Question 4

What is the Problem transformation approach?

Answer 4

Transforms a multi-label problem into one or many single-label classification problems
- Uses a single label classification algorithm in each problem

Question 5

What is the algorithm adaptation approach?

Answer 5

Adapts a single-label classification algorithm for multi-label problems

Question 6

What is Binary Relevance method (This is a Problem Transformation method)?

Answer 6

Creates one single-label dataset for each label in label set. All datasets share the same features but different class labels.

Question 7

What is the pro and cons of Binary Relevance method?

Answer 7

Pro:
simplicity

Cons:
Requires many runs of a classification algorithm.
Ignores interactions among class labels.

Question 8

How does the the Power Set method work (Problem Transformation method)?

Answer 8

Creates a new label for each combination of labels in the original dataset. Each instance is assigned one label.

Question 9

What is the pro and con of Power Set method?

Answer 9

Pro:
Only one run of a standard multi-class algorithm is needed. So it is faster than binary relevance.

Con:
Leads to datasets with many compound class labels, with very few examples for many of those labels.

Question 10

What Problem Transformation method is the most advanced?

Answer 10

Classifier Chain method

Question 11

How does Classifier Chain method work (Problem Transformation method)?

Answer 11

First, create one dataset for each label

Second, specify an ordered list (chain) of class labels

Third, extend dataset to contain not only original features but also features representing the labels predicted by previous classifiers in the chain

Fourth, use a classification algorithm to predict labels based on the chain order, propagating predicted labels along the chain

Question 12

What is the pro and con of Classifier Chain method?

Answer 12

Pro:
Label interaction

Con:
Computationally expensive