Session 6

Question 1

Why use clustering?

Answer 1

1. Data exploration: “looking for “interesting patterns” without prescribing any specific interpretation”.
2. Information reduction.
3. Verification.

Question 2

Clustering in the CRISP-DM cycle

Answer 2

Mostly:
• Data understanding - exploration
• Data preparation - preprocessing, reducing dimensionality
• Modelling

Question 3

How to interpret clusters?

Characteristically.

Answer 3

Characteristically.

• Interpreting clusters by looking at its members
OR
• Interpreting clusters by looking at a typical cluster member or typical characteristic(s).

Question 4

How to interpret clusters?

Differentially.

Answer 4

Differentially.

• What differentiates Cluster X from Cluster Y?

• Supervised learning approach.
‣ Each data point has a new label - cluster ID.
‣ Predictive modelling with cluster ID as a target variable.

Question 5

How would that work? Supervised learning approach - Differentially?

Answer 5

Set up a classification task:
1) a k-class task or
2) binary classification
• Ensure intelligibility - be able to get
the classifier definition

Question 6

Cluster validity

How good is a given clustering?

Internal criteria

Answer 6

• uses the internal information of the process to evaluate the quality of clustering and to what extent it fits your data.
• Compactness and isolation.

• Can be used to specify the optimal number of clusters.
Indices are often (not always) method-dependent.

Question 7

Cluster validity

How good is a given clustering?

External criteria

Answer 7

• Compare the created clusters to some external reference:
‣ Experts opinion, existing theory.
‣ External variables/groupings.
‣ Labels generated by a different clustering method.

• Similarity between the two clusterings.
• Can be used to find a suitable clustering algorithm for a given data set.

Question 8

Steps in a typical cluster analysis:

Answer 8

A. Collect the data to use for clustering. Preprocess if needed.
B. Select the variables.
C. Select distance measure.
D. Select clustering method.
E. Experiment with different sets of variables/measures/methods.
F. Determine validity of the selected solution.
G. Interpret the results.

Experiment! - difficult to anticipate what combinations of variables, similarity measures, clustering methods will lead to interesting results.

Question 9

Random Forest model to…

Answer 9

assess the explanatory power of each variable -> variable importance

Question 10

Where else can we find clustering applications?

Answer 10

• Identifying fake news
• Document analysis (organising the information)
• Weather analysis (e.g. understanding meteorological patterns)
• Genetics