Cluster Analysis

Question 1

Cluster analysis is mainly used to aggregate of CASES, not FIELDS:

Answer 1

TRUE, some programs (like SPSS) allows the user to group also FIELD but this is far from being a standard use. CLUSTER is used to create groups of OBJECTS, cases, not variables.

Question 2

TWO STEP cluster is able to balance, automatically, the size of resulting clusters

Answer 2

FALSE. Sizes we get for different groups depend on the natural dispersion of groups according to our settings (variables, distance metrics,….) and CANNOT be balanced automatically…

Question 3

Both Euclidean Distance or Maximum Likelihood can be used in 2-Step Cluster as distance measurements, no matter the type / measurement of clustering variables we are using

Answer 3

FALSE. When we mix categorical and metric variables in a TWO STEP procedure, we SHOULD used Log-Likelihood.

Question 4

We need homogeneous sets of data for hierarchical clusters (scale or categorical, but not mixed)

Answer 4

TRUE. At least with SPSS we can only use a type of variables because the measure of distance we may select depends on the type of variables we have (scale, counts or binary).

Question 5

A dendrogram can be used to identify outliers

Answer 5

TRUE. Outlier cases would appear in the dedrogram as cases/objects far from any others (such ME in the MATCHING example for the class)

Question 6

A dendrogram does not provide any support for selecting the proper number of clusters

Answer 6

FALSE. We CAN explore the structure of clustering in order to get a visual idea about the “ideal” number of clusters in terms of homogeneity. For instance, a dendrogram may shows that a solution with THREE clusters entails to MIX two different groups that are far from each other thus suggesting that FOUR clusters is more appropriate.

Question 7

The choice of the distance measure in hierarchical clusters depends, basically, on the number of variables

Answer 7

FALSE. It depends on the measurement of variables.

Question 8

A Two-Step cluster produce solutions based on mixtures of continuous and categorical variables

Answer 8

TRUE. This is in fact one of its benefits compared to other algorithms.

Question 9

One of the problems of Two-Step is its inherent inability to handle outliers

Answer 9

FALSE. In fact, this is another advantage of Two-Step; in fact, the algorithm is able to automatically find and filter outlier cases (see class doc for more info on this)

Question 10

The standardization of variables is a quite common as a preliminary step in cluster exercises

Answer 10

TRUE. If we don’t standardize, the size of each variable will impact in the distance metric and thus in the cluster result (remember what we saw in a class example about the importance of AGE compared to number of CHILDREN.

Question 11

A balanced distribution of cluster sizes is COMPULSORY

Answer 11

FALSE. IT IS NOT COMPULSORY even if sometimes, balanced sizes for different clusters will be good news.

Question 12

A hierarchical cluster is normally nonsense when having thousands of records to cluster

Answer 12

TRUE. Hierarchical is normally used when we want to track aggregation “path” or “process” ONLY for some selected cases of special interest among a limited number of other cases.

Question 13

Technical quality of a CLUSTER output can be evaluated with the Silhouette measure

Answer 13

TRUE. Even if this is only about a technical assessment, it is true. It measures the compactness of each group and the separation between groups.

Question 14

The size of clusters is NOT a relevant matter at all

Answer 14

FALSE. It is normally a relevant matter. We normally seek for large groups where we can profit any action so very small groups are not normally good news.

Question 15

A cluster exercise is mainly a PREDICTIVE analytical exercise

Answer 15

NOT at all. We are not PREDICTING anything. We don’t have any “target”. CLUSTER is an UNSUPERVISED technique, just about an exploratory analysis.

Question 16

Normally, a good cluster solution is shaped with a LARGE number of clustering variables (not less than 15)

Answer 16

FALSE. Normally it’s the other way around. A Cluster solution with lots of “dimensions” or input variables is a bit impractical so we typically prefer a solution depicted by a limited number of input fields.

Question 17

A good cluster produces clusters with low SEPARATION and high COHESION

Answer 17

FALSE. A perfect group structure means HIGH cohesion (within each group) and HIGH separation (between groups).

Question 18

The SILHOUETTE measure in a Cluster output is mainly used to evaluate the optimum number of clusters to get.

Answer 18

FALSE. It is used to evaluate the overall technical quality of our result in terms of SEPARATION and COHESION and this technical value depends on every cluster setting (selected variables, measure of distance,…) and not ONLY on the number of groups. It is true that the NUMBER of clusters will also impact this technical quality but SILHOUETTE is NOT the standard tool to select the NUMBER of clusters.

Question 19

TWO STEP Cluster is called that way because the initial cluster solution is internally validated in a holdout sample during a second stage

Answer 19

FALSE. The name “two-step” comes from the idea of a double – stage when producing clusters (see technical annex in the class doc).

Question 20

A field/variable can be very relevant to define/ distinguish a SPECIFIC CLUSTER without being of great importance for the solution as a whole

Answer 20

TRUE, this is true. Imagine a group solution with 4 groups of consumers (A, B, C & D). Imagine that two groups (Group C & Group D) are BOTH different from A and B in terms of age, average expenditure, and recency …. BUT similar between them and the only difference is that group D is about consumers living in rural areas. This “rural areas” may be ONLY of marginal interest to define group D, and not to understand A,B or C.

Question 21

TWO STEP Cluster is called this way because it mixes two different algorithms to produce a final solution

Answer 21

TRUE. See technical annex in the class doc for more info.

Question 22

EUCLIDEAN distance can be used when using TWO – STEP under certain assumptions

Answer 22

TRUE. When we only use scale variable we can still use Euclidean distance.

Question 23

The selection of the clustering VARIABLES highly conditions the cluster solution we get

Answer 23

TRUE, TOTALLY. Groups we will get are completely conditional of the variables we use to understand similarity or distance.

Question 24

The measure of DISTANCE does not have any impact on the cluster solution

Answer 24

FALSE. It has an impact, and not only because of technical reasons BUT also because, from a conceptual point of view, DISTANCE is the way we understand dissimilarity so by changing our notion of difference Vs similitude we change our understanding of groups.

Question 25

A Two Step cluster - Outliers will be detected automatically

Answer 25

TRUE. The option “Outlier treatment - use noise handling” in the second screenshot will make SPSS to filter –out outlier cases automatically (see class document about Two – Step for further details)

Question 26

We are using log-likelihood distance but Euclidean would have been a better choice

Answer 26

FALSE. Euclidean is NOT an option when we mix categorical and scale variables.

Question 27

Considering that we have 10.000 customers in our dataset, a hierarchical cluster would have been a better option that TWO STEPS

Answer 27

FALSE. Precisely because we have thousands of records, a hierarchical cluster doesn’t look a good option.

Question 28

It is especially useful to find groups in a multivariate context

Answer 28

Yes. Finding groups without a Cluster algorithm is only feasible in univariate / bivariate contexts but, in the presence of several features we normally need a technical algorithm in order to find compact and well separated groups.

Question 29

Its supervised, in the sense that we always will consider a clear leading variable to define or discover groups

Answer 29

No. “Supervised” means that we have a TARGET variable to be explained or predicted. In a cluster exercise we may have several inputs (clustering variables) but not a TARGET variable.

Question 30

Can only be used with scale/continuous variables in our dataset

Answer 30

No. there are distance measures able to process either categorial and scale variables and even a mix of them

Question 31

A cluster exercise is more about exploring grouping structure, that about confirming the existence of certain groups

Answer 31

Yes. It is a exploratory tool and, in fact, we normally get, inspect and compare several cluster solution before making a decision on the one that fits our needs.

Question 32

Cluster is of great interest in the context of Market Segmentation

Answer 32

Yes. It is not only used for Market segmentation but its use in this specific context is even considered “a must”.

Question 33

There is no way of evaluating a cluster solution from a technical point of view

Answer 33

False. Even if technical quality is not the most important feature to be evaluated in a cluster solution, there are several metrics that we can use in order to asses this technical quality (Silhouette, for instance).

Question 34

Cluster can be mainly used to reduce the number of variables in a dataset

Answer 34

No. This technique is not about dimensions reduction; it explores homogeneity across individuals or objects (rows in a dataset) not across variables (columns in a dataset).

Question 35

We don’t need a target variable to run a Cluster analysis

Answer 35

Right. This is not a supervised technique.

Question 36

To some extent, a good or bad cluster result is not only about technical features but also depends on the context of our analysis

Answer 36

Exactly. A cluster solution should be always aligned with our expectations and be useful in terms of action even if it is not the optimum one in technical terms.

Question 37

The number of clusters to get is usually clear from the begining of a cluster exercise

Answer 37

Not at all. Sometimes we may have a preliminar idea about the optimum number we would like to get, but not about the number we will finally get when we use a given algorithm.

Question 38

It is useful to use as many variables as you can at the same time to improve clusters definition

Answer 38

FALSE. Including LOTS OF VARIABLES is normally a bad strategy and, of course, does not implies a better solution.

Question 39

We don’t necessarily need to standardize metric variables IF ALL OF THEM of them are comparable because they are expressed in similar units / scales

Answer 39

TRUE. Standardization is very frequent BUT it is only needed when scale variables are expressed in different units.

Question 40

Euclidean distance is a standard in the presence of metric variables or binary variables

Answer 40

TRUE. It’s only main limitation is for handling categorical variables.

Question 41

It is useful and common to drop or exclude atypical or anomalous cases to find a clearer Cluster solution

Answer 41

TRUE. This is mentioned in the class doc and it is always a GOOD advice for every single analytical procedure you may want to use. Concentrate in what is “normal” without paying attention to odd cases.

Question 42

There is a variety of distance measures we can use that can produce very different clusters results

Answer 42

TRUE. Distance affects results (among other settings).

Question 43

The distance measure to choose depends critically in the NUMBER of variables we need to use for our clustering exercise

Answer 43

FALSE. Depends on the measurement of variables

Question 44

We run a hierarchical cluster when we want to find a clustering solution for thousands of records

Answer 44

FALSE. We run a hierarchical for a limited number of objects when we want to explore the agglomeration schedule for some objects/subjects of interest.

Question 45

We can use a hierarchical cluster combining metric and categorical variables

Answer 45

FALSE. We can only use a type of measurement, not a mix of both.

Question 46

We are usually interested in the agglomeration process of a hierarchical clustering solution

Answer 46

TRUE. This is exactly the aim when we use a hierarchical algorithm

Question 47

If we want to explore and understand simmilarities between our 25 competitors in the market, a hierarchical cluster looks a good solution

Answer 47

TRUE. This would be good example. Our interest is to position our brand compared to others, so we may need to inspect distances / proximities brand by brand.

Question 48

A hierarchical cluster always uses an agglomeration procedure

Answer 48

FALSE. The algorithm might be divisive (check the illustration in the second class-doc)

Question 49

A hierarchical cluster using “N” observations will always offer a range of cluster’s solutions from “N” to “1” clusters

Answer 49

TRUE. At the beginning of the procedure, each case is a group (N groups) and the process always ended with every case in the same group (1 group)

Question 50

A dendrogram graph can be used to:

Answer 50

Can be used to detect outliers

Question 51

The Distance Matrix can be used to:

Answer 51

• Can be used to explore simmilarity between individuals/objects
• Shows the multivariate distance between objects/individuals
• Is interesting, and essential to run the agglomeration procedure

Question 52

We could use a crosstabs to check gender distribution across clusters

Answer 52

TRUE. Both variables are categorical

Question 53

We could use a Box Plot Graph to explore age (in years) across clusters

Answer 53

TRUE. Age (if it is measured in years) is a metric variable and cluster membership is categorical, so we may use a Box-Plot.

Question 54

We would use ANOVA to check statistical differences in age (in years) across clusters if we have more than two clusters in our solution

Answer 54

TRUE. This is the test for the combination “metric variable & categorical variable” when the categorical variable has more than two categories. If we only have TWO clusters we could use a t-test.

Question 55

We cannot use a linear correlation analysis because our cluster membership variable is not metric

Answer 55

TRUE. We can only use it for two METRIC variables.

Question 56

TSC could be a good option if we have thousands of records to cluster

Answer 56

TRUE. Hierarchical is not an option for thousands of records.

Question 57

TSC output shows a dendrogram if needed

Answer 57

FALSE. It is not an option and, besides, it would be impractical with thousands of records in the dataset.

Question 58

TSC helps us to automatically identify the “optimum” number of clusters (from a technical point of view)

Answer 58

TRUE. Yes, at least in SPSS this is an option.

Question 59

The name TWO-STEP is because TSC combines internally two different procedures in two different stages.

Answer 59

TRUE. Check technical annex in the class doc if you are interested in details.

Question 60

Two Step can only produce solutions for continuous variables.

Answer 60

FALSE. IN fact, this is one of the advantages of TWO-SETP (the ability to deal with a mix of continuous and categorical variables)

Question 61

There is commonly a penalty to the cluster quality for assuming a certain probability distribution for continuous and categorical

Answer 61

TRUE. Check the class doc (disadvantage number 3, page 14).

Question 62

TSC is part of the hierarchical clustering algorithms family

Answer 62

FALSE. It is not hierarchical.

Question 63

TSC can NOT handle outliers automatically

Answer 63

FALSE. In fact, this is one of the advantages of TWO-STEP.

Question 64

We should better standardize our METRIC clustering variables using TSC in SPSS if they are not comparable in units.

Answer 64

TRUE. If we want to avoid an impact of units in the cluster output.

Question 65

If you want to explore different travellers profiles from your large database of trips reservations during the last year, Which algorithm would you use?

Answer 65

Two Step Cluster looks good. If we have a large dataset of transactions and, besides, one assume that we have to deal with a combination of metric and not metric variables, a hierarchical cluster is not a good option.