Store segmentation progresso

Multivariate Analysis
Segmenting Stores in Soup Case Study
D3M
This is the store demo file

Objective
 Segment the 2000 IRI stores into smaller groups
 Interpret the segments you created
 Compute the price elasticity for each segment and
discuss the pricing strategy that Progresso should
pursue to maximize profits
 State the assumptions used in deriving optimal prices for
profit maximization
 Discuss the practicality of your recommended pricing
strategy

Approach
 Questions you should ask
 Segmentation based on what??
 How many segments??
 Always start by summarizing variables in your
data and understanding the basic relationships
 Understand the correlation b/w variables –store
demographics & market shares
 These are what we will use for segmentation

As usual, start by summarizing the data

Several of the
Demographic
variables are Highly
Correlated

Correlations of Market Shares Across 2000 Stores
What can we learn about Progresso’s Competitors from just correlations?

Campbell does well in Midwest
& South
Progresso is strong in East,
followed by West

Segmentation of IRI Stores
D3M

Factor & Cluster Analysis
Learning Objectives
 Unsupervised Learning Methods
 Principle component, Factor Analysis, & Clustering
 Objective is Dimension Reduction
 Reduce the number of collinear variables (PCA/Factor)
 Group your rows (e.g. customers, markets, counties): Cluster Analysis
Additional Learning Resources
 MIT Open Courses Lecture 11 & 14
 Data Mining Class at U of Chicago (Lecture notes 7 & 8)
 Stanford course on Machine Learning: Watch Lecture 10 on
“Unsupervised Learning”

Note the Difference between Cluster and PCA/Factor analysis
V1 V2 V3 V4 V5 V20…..
Cluster
Analysis
(Group Subjects)
Factor
Analysis
(Group Variables)
Data

Variable Reduction Techniques
You are working with columns here
We will look at 2 Techniques
 Principle Component Analysis
 Factor Analysis

PCA/Factor Analysis
 Our demographic variables are highly correlated
 If we were to use these in a Regression model for example, we will high
multicollinearity
 A useful technique for reducing the number of variables is
Principle Component Analysis (PCA) & Factor Analysis
 PCA/Factor analysis is able to summarize the information
contained in a larger number of variables into a smaller number
of ‘factors’ without significant loss of information
 Widely used technique in a variety of fields ranging from
Psychometrics to analysis of unstructured data like text or
images

If we use 3 components, we capture approximately 84% of information
contained in the 10 demographics
Eigenvalues of a matrix are also
called characteristic roots and
represents the variance accounted
for by a linear combination of the
variables. Usually # of components
to use is Eigenvalue greater than 1.
In our case its 3
Principle Component Analysis

Look for large positive or negative numbers for
each factor. See the corresponding variable
names to interpret the underlying ‘factor’
These are called factor “loadings”. Measures the correlation between each demographic
and the underlying “factor”. Our Job to Interpret and put a label to these.
Factor Analysis
Using 3 “factors” instead of 10
demographics, we capture approx.
84% of the information.

What do these techniques do?
 Take a large number of variables
that are highly correlated & create
new variables
 New variables (components or
factors) are linear combinations of
our current variables
 Goal is to retain most of the
variability (information) in the data
 Reduce the dimension of the
problem with little loss of
information
 Newly created variables are
orthogonal (no correlation)
Note: Our current application of 10 demographic variables is
quite trivial. We will see larger problems where these methods
are more useful
These are the
new variables
in our data.
Our job is to
interpret
them. The
new variables
(factors) are
standardized
and
uncorrelated.
We can use
them further
for other
analysis, for
example
Segmentation
of stores in
our data.

Examine the Factor Scores
The new variables (Factors) have a mean of 0 and Std of 1.
They are orthogonal to each other (zero correlation)

Cluster of Variable Algorithm
We can use Median
Income, % Kids 18,
and % Black. These
3 variables will be
representative of
other demographics
in its cluster

Cluster Analysis
Segmentation of IRI Stores
D3M

Now we are interested in grouping rows (Stores in our case)
V1 V2 V3 V4 V5 V20…..
Cluster
Analysis
(Group Subjects)
Factor
Analysis
(Group Variables)
Data

21
Cluster Analysis
Cluster analysis is a technique used
to identify groups of ‘similar’
customers in a market (i.e., market
segmentation).
Cluster analysis encompasses a
number of different algorithms and
methods for grouping objects of
similar kind into categories.

22
General question: how to organize observed
data into meaningful structures
• Examples:
o In food stores items of similar nature, such as
different types of meat or vegetables are displayed in
the same or nearby locations.
o Biologists have to organize the different species of
animals-- man belongs to the primates, the
mammals, the amniotes, the vertebrates, and the
animals.
o In medicine, clustering diseases, cures for diseases,
or symptoms of diseases can lead to very useful
taxonomies.
o In the field of psychiatry, the correct diagnosis of
clusters of symptoms such as paranoia,
schizophrenia, etc. is essential for successful
therapy.
o Collaborative filtering & Recommendation systems

23
Cluster Analysis
Cluster analysis works on the principle of maximizing the between-
cluster variance while minimizing the within cluster variance
Methods: Hierarchical & K-mean Clustering

Clustering Methods
 Hierarchical clustering is an iterative process that starts with
each observation in its own cluster. At each stage, the
algorithm combines two clusters that are closest together. At
the final stage, all observations are in one cluster.
 Useful for small data sets, takes a long time for large tables.
24
 K-means clustering starts with a known number of clusters, k. The
algorithm picks k cluster seed points, then assigns each observation
to a cluster. It then replaces the cluster seeds with the cluster
means and repeats until the clusters stabilize.
 Works well with large data sets

Hierarchical Clustering of Stores
Questions to Ask: Clustering based on what? How Many Segments?

Exercise
 Conduct a Hierarchical cluster analysis based on
 Saved Factor Scores & Market Shares of Brands
 To keep things manageable, lets use a 5-segment solution
 Interpret the clusters based on
 Median Income, % Kids Under 18, % White, & Market Shares
 What segment has the highest appeal for Progresso?
 Save the cluster membership and merge file with Transaction
data
 Redo the regression analysis and analyze the own & cross-price elasticity in
each segment
 Suggest an optimal pricing strategy for Progresso for each segment
 Discuss practical considerations in using such segmentation/pricing scheme

Store segmentation progresso

More Related Content

What's hot (19)

Similar to Store segmentation progresso (20)

More from veesingh (11)

Recently uploaded (20)

Store segmentation progresso