5 - Clustering, Segmenting, and Cutting Through the Noise

Question 1

What is the primary goal of the Fraud Department at Shu Financial?

Answer 1

To protect the company from losses and ensure customers’ private information is secure

Question 2

What are the two types of identity fraud?

Answer 2

• Application fraud
• Account takeover

Question 3

What is a supervised machine learning model?

Answer 3

A model trained on a target variable, such as the probability of an application being fraudulent

Question 4

True or False: The Fraud Department is considered a profit center for Shu Financial.

Answer 4

False

Question 5

What is the focus of unsupervised machine learning in the context of fraud detection?

Answer 5

To find patterns and groupings among fraudulent applications without a predefined target variable

Question 6

What is principal components analysis (PCA)?

Answer 6

A dimensionality-reduction method that reduces the number of variables while retaining important signals

Question 7

What does ‘dimensionality reduction’ refer to?

Answer 7

The process of reducing the number of columns in a dataset while retaining important information

Question 8

What types of variables does PCA work best with?

Answer 8

• Continuous variables
• Ordinal variables

Question 9

What is the first preprocessing step before conducting PCA?

Answer 9

Eliminating variables that have limited incremental information

Question 10

What is normalization in the context of data standardization?

Answer 10

Adjusting variables to have a mean of 0 and a standard deviation of 1

Question 11

What is the relationship between principal components?

Answer 11

Each component is perpendicular to the previous components, meaning there is no correlation

Question 12

What is cluster profiling?

Answer 12

The process of generating unique descriptions for each cluster using the input variables

Question 13

What should be done if the clusters are not well separated?

Answer 13

Investigate the features to describe what the clusters represent

Question 14

What is the role of the business unit in the data science process?

Answer 14

To provide domain-specific knowledge that helps interpret principal components and clusters

Question 15

Fill in the blank: The second principal component is computed by finding what combination of input variables explains the _______.

Answer 15

remaining information not explained by the first principal component

Question 16

What does the data science team plan to do after identifying different groups of fraudulent applications?

Answer 16

Build separate models to target the different clusters

Question 17

What is t-SNE?

Answer 17

A dimensionality reduction method often used when separating clusters using linear assumptions is difficult

Question 18

What happens if too much critical information is discarded during dimensionality reduction?

Answer 18

It can hinder the analysis and lead to less effective models

Question 19

What is the importance of the partnership between the business unit and the data science team?

Answer 19

It ensures that the data science products meet business needs, avoiding wasted time and resources

Question 20

What does the first principal component typically represent?

Answer 20

The combination of variables that explains the most information in the dataset

Question 21

True or False: Categorical variables are very useful in principal components analysis.

Answer 21

False

Question 22

What is the process of profiling clusters based on?

Answer 22

Labeling clusters based on a few variables that are the most extreme in that specific cluster

Question 23

Who are the clients in the collaborative profiling process?

Answer 23

Steve and experts in the Fraud Department

Question 24

What is the limitation of manually labeling clusters when the number of clusters is very large?

Answer 24

A manual process of labeling clusters is not feasible

Question 25

What can the data science team track when exploring different clustering algorithms?

Answer 25

Number of clusters identified, percentage of the population in each cluster, and cluster descriptions

Question 26

What types of characteristics might be used to profile clusters?

Answer 26

* Demographics (age, sex, location) * Transaction types (gas stations, ATMs) * Patterns in online transactions

Question 27

Who may be called in to review the clusters for assessability?

Answer 27

Subject-matter experts

Question 28

What should Steve discuss with operational teams regarding modeling results?

Answer 28

How they plan to use the model output and their capacity limitations

Question 29

What is a potential consequence of identifying too many clusters?

Answer 29

Information may not be useful and could exceed the capacity of the team

Question 30

What might investigators do if they find excessive cluster information?

Answer 30

Begin to ignore the modeling output

Question 31

What can streamline processes for application fraud types?

Answer 31

Developing streamlined processes for five different types of application fraud

Question 32

What is one benefit of data exploration in predictive modeling?

Answer 32

Developing new features that can be used to improve predictive models

Question 33

Fill in the blank: The data science team could examine the details of the first few principal components to see if they represent features that are not already explicitly in the _______.

Answer 33

predictive model

Question 34

What should the modeling align with to be effective?

Answer 34

Operational capacity

Question 35

True or False: A major opportunity loss for the organization can occur if the team exceeds their capacity to handle cluster information.

Answer 35

True

Question 36

What is unsupervised machine learning?

Answer 36

A broader category of machine learning that includes clustering as one of its common applications.

Question 37

What is the main goal of clustering in the context of fraud analysis?

Answer 37

To group similar fraud cases together and separate dissimilar fraud cases.

Question 38

List some common clustering algorithms.

Answer 38

* Gaussian mixed models * Expectation-maximization models * Latent Dirichlet allocation methods * Fuzzy clustering algorithms

Question 39

What distinguishes exclusive clusters from fuzzy clusters?

Answer 39

Exclusive clusters assign customers to only one group, while fuzzy clusters allow customers to belong to multiple clusters with varying probabilities.

Question 40

What is hierarchical clustering?

Answer 40

A type of clustering where smaller clusters belong to larger clusters, often visualized like branches of trees.

Question 41

What is partitional clustering?

Answer 41

A clustering method that does not involve hierarchy and focuses on dividing data into distinct groups.

Question 42

What preprocessing step is commonly performed before clustering?

Answer 42

Standardizing features so that they have similar dimensions.

Question 43

What is K-means clustering?

Answer 43

A partitioning-based clustering method that assigns observations to clusters based on their distance to centroids.

Question 44

How does K-means determine the optimal number of clusters?

Answer 44

By testing a range of values for K and selecting the number that explains the most variance in the data.

Question 45

What is a centroid in K-means clustering?

Answer 45

The center point of a cluster, calculated as the mean of all points in that cluster.

Question 46

True or False: K-means clustering can be heavily influenced by the initial starting points.

Answer 46

True.

Question 47

What is fuzzy K-means clustering?

Answer 47

A variation of K-means where customers can belong to more than one cluster, with probabilities based on distance from centroids.

Question 48

What is the difference between agglomerative and divisive clustering?

Answer 48

* Agglomerative: Bottom-up approach merging clusters * Divisive: Top-down approach splitting clusters

Question 49

What is the significance of distance measurement in clustering?

Answer 49

It impacts how clusters are formed and merged, affecting the final cluster shapes.

Question 50

What are some common types of fraud that may be clustered?

Answer 50

* Fraudulent applications * Account takeovers * Card thefts * Illegal use of lost cards * Skimming

Question 51

What is the purpose of cluster analysis in operational settings?

Answer 51

To interpret data and improve decision-making in fraud detection and prevention.

Question 52

What was one outcome of the cluster analysis performed by Steve's team?

Answer 52

The development of behavioral-based predictive models that target specific features of different clusters.

Question 53

What is the role of principal components analysis in clustering?

Answer 53

To reduce dimensionality and identify the most significant components contributing to variance in the data.

Question 54

What is a key question to ask regarding the clustering algorithm used?

Answer 54

What metric was used for the distance in the clustering process?

Question 55

Fill in the blank: The final number of clusters in Steve's project was set at _______.

Answer 55

six.

Question 56

What is dimensionality reduction?

Answer 56

A technique used to reduce the number of features in a dataset while preserving its variance.

Question 57

What should be considered when choosing a method for dimensionality reduction?

Answer 57

The method used and the rationale behind its selection.

Question 58

What is a key question regarding variance in dimensionality reduction?

Answer 58

How much variance is explained by the different components?

Question 59

What is an important consideration when focusing on components in dimensionality reduction?

Answer 59

Whether reducing focus to just a few components leads to loss of too much information.

Question 60

How should components be interpreted in dimensionality reduction?

Answer 60

By analyzing what the first, second, and third components represent.

Question 61

What is a key characteristic of clustering algorithms?

Answer 61

Whether the clustering is exclusive.

Question 62

What must be specified when discussing clustering algorithms?

Answer 62

Which clustering algorithm was used.

Question 63

What metric is important in clustering algorithms?

Answer 63

The distance metric and the linkage criteria between clusters.

Question 64

Why is it important to discuss distance and linkage in clustering?

Answer 64

To understand the criteria used for clustering.